How to Plan a Data Story That Other Publishers Can Verify and Use
If you want another publisher to cite, adapt, or link to a data story, plan for that editor to check the finding…
Categories
How to Measure SEO and GEO Results
What SEO and GEO measure
A business may receive organic-search clicks without knowing how often customers encounter it in AI-generated answers. The reverse also happens: an AI answer may mention or recommend a business without sending visitors to its website. Neither signal proves that the business gained qualified leads, sales, or revenue.
SEO measurement tracks visibility in organic search results, visits from those results, and actions recorded after a visitor arrives. GEO, short for generative engine optimization, measures how a business appears in generative search and answer products. It records what the answer says: whether it mentions the business, recommends it for a stated need, or links to one of its pages.
A displayed link shows that the answer product presented a source. A recommendation shows that the answer presents the business as a fit for the question. A mention may do neither; for example, the business may appear in a comparison list. None proves that a customer visited, converted, or generated revenue.
Keep the evidence layers separate
Reliable measurement separates four kinds of evidence:
- Technical access: whether a web request retrieves a page.
- Search visibility: whether a search engine shows a page for a query.
- Answer visibility: what an AI answer says, recommends, and links to for a prompt.
- Business outcomes: the leads, purchases, calls, sign-ups, and other results recorded by the business.
For example, on November 11, 2026, Article Lab received HTTP 200 responses for the resources page and the sitemap audit page. An HTTP 200 response means the server successfully returned the requested page to an ordinary web request at that time.
That result does not show that Google or another answer product fetched, indexed, cited, or sent customers to either page. Each claim needs evidence from its own layer. Search Console supports claims about Google Search impressions and clicks. Saved answer outputs support claims about what a measured answer product displayed. Analytics, call tracking, and sales records support claims about business results.
AI-answer behavior depends on the product
Generative search products do not retrieve, compose, cite, or display answers in the same way. Record the product and prompt. Also note the location, date, account state, and other conditions that shaped the observed answer.
In documentation updated December 10, 2025, Google states that AI Overviews and AI Mode may use query fan-out: related searches across subtopics and sources that help build an answer. Google also states that these features may use different models and techniques, so answers and linked sources vary. These statements apply to Google Search’s AI features, not every answer product. Google’s AI features documentation also states that Search Console includes traffic from these features in Web search reporting.
A page cited in one Google AI Overview may not appear for a closely related prompt, on another date, or in a different answer product. A score based on such observations describes only the measured sample unless reviewers can inspect the prompts, dates, answer outputs, and scoring rules.
Measure visibility without overstating impact
Keep the underlying observations, repeat comparable queries under comparable conditions, and state uncertainty plainly. Connect search and answer visibility to business outcomes only when tracking data supports that connection.
Key takeaways
- SEO records organic search visibility, along with visits and the actions people take after visiting.
- GEO records how a business appears in AI-generated answers. It tracks mentions, recommendations, and displayed links.
- Technical access, search visibility, answer content, and revenue are separate evidence layers.
- A visibility score describes what appeared in a defined sample. It does not prove revenue impact.
A necessary counterargument
A business does not need perfect measurement before improving pages, products, or customer information. Clear descriptions, accurate pricing and service details, crawlable pages, and useful original material support customer discovery and conventional search performance. An AI mention or a successful page request still does not prove that those changes produced revenue.
Choose the business outcome
Start with the decision the business needs to make. A service firm that needs more qualified inquiries should focus on suitable prospects who contact the firm, not brand mentions alone. An ecommerce store that needs better product discovery should focus on completed orders and net revenue.
Define the business outcome before choosing a metric. An online store might measure completed orders and net revenue, using a written rule for returns, cancellations, tax, and shipping. A service firm might measure qualified inquiries and closed revenue, with written criteria for a qualified inquiry.
Keep completed work in a separate record. Publishing a page, correcting a product fact, changing a price, or adding structured data are inputs. Record those actions beside the results, but do not treat them as proof that search visibility, answer visibility, leads, or revenue improved.
Keep measures in layers
A small measurement map prevents unlike evidence from becoming one ambiguous visibility score.
| Measure | What it establishes | What it leaves unanswered |
|---|---|---|
| Access and indexing | Whether a page responds to a request and whether Google reports it as indexed | Whether the page appeared for a buyer’s search or persuaded a visitor |
| Search performance | Whether Google reported impressions and clicks for a page or query | Whether an AI-generated answer caused a visit or sale |
| Answer observations | What an answer product displayed for a recorded question | Whether users saw the answer, trusted it, or acted on it |
| Business outcomes | Whether tracked visits, qualified leads, orders, and revenue occurred | Which discovery source caused the outcome |
Access and indexing answer different questions. A page that returns HTTP 200 responded successfully to the request that checked it. That result does not show that Google fetched, indexed, or ranked the page. Google’s Page indexing report shows Google’s indexing status for pages on a site. An indexed page still may not appear for a buyer query.
Search performance records activity in Google Search results. Google Search Console’s Performance report lists clicks and impressions and supports analysis by page and query. Google states that traffic from AI features, including AI Overviews and AI Mode, is included in overall Web search traffic. Search Console does not provide a separate report for that traffic (Google Search documentation). Do not describe a Web search total as AI-feature traffic.
Answer observations describe what a specific AI answer product showed for a specific question. Record whether the answer mentioned the business, recommended it, linked to a source, or made a factual claim. A mention does not show a recommendation. A displayed link does not show a click. Check whether a cited page supports the claim attached to it.
Business outcomes record what happened after discovery: identifiable visits, qualified leads, completed purchases, or revenue. State how each outcome connects to a source and what the tracking misses. Google Analytics’ Traffic acquisition report groups recorded sessions by source and channel, showing where visits came from without capturing every influence that led a buyer to visit or purchase.
Set the comparison before measuring
Choose the reporting period and comparison period before reviewing the figures. Use comparable windows, such as the same number of days and equivalent days of the week. Record promotions, price changes, product availability, and launches. Record seasonality and site migrations too.
These notes provide context; they do not prove why a number changed. If organic clicks and orders rise during a promotion, the figures do not separate the promotion’s effect from changes in rankings, demand, or tracking.
For answer-product checks, keep the question panel stable across comparisons. Use the same prompts, product and mode, locale, and measurement dates. Save the full answer and displayed sources. Separate branded questions, which name the company or product, from unbranded questions, which describe a need without naming a supplier. If the panel changes, report the new question mix. Counts and percentages can move solely because the questions changed.
One recorded answer is evidence of one run under recorded conditions. Repeating the check shows whether the answer changes. A small question panel represents only a narrow set of possible searches, prompts, and users.
Counterargument: one score looks simpler
A single score makes reporting easier when leaders want one trend line. It also hides the decision behind the number. A score that combines indexed pages, impressions, AI mentions, visits, and revenue can rise while qualified leads fall. It can also fall while revenue rises.
Report the layers side by side. Decision-makers can then see whether the constraint lies in technical access, search exposure, answer representation, conversion, or missing measurement.
Check access before visibility
An HTTP response records what a specific request received and when, but provides no evidence that Google or an AI provider fetched the page, indexed it, or selected it to support an answer. Keep technical access checks, provider indexing evidence, and observed search or answer visibility separate.
For each important URL, start with its status code and redirect chain. Then review the page-level signals: the canonical tag and robots directives. Finish by checking the sitemap entry and important text in the response HTML. A canonical tag identifies the URL that the site wants search engines to treat as the main version of a page. Page-level robots directives give search engines instructions for a specific page. robots.txt gives crawl instructions to crawlers that choose to follow it. A noindex meta tag or X-Robots-Tag HTTP header tells search engines not to index a page or response.
When important content appears in a browser but not in the response HTML, compare the same page sections. JavaScript may add content after the server sends the initial response. Logged-in views, consent choices, personalization, and comment forms also create browser-only differences.
Google states that a page must be indexed and eligible to appear with a snippet in Google Search before it is eligible as a supporting link in AI Overviews or AI Mode. That condition does not mean Google will select, cite, or recommend the page. It applies to Google Search features, not other answer products. See Google’s documentation for AI features in Search.
For Google-specific evidence, use URL Inspection and Search Console’s indexing reports. These tools report Google’s information about a URL and its indexing state. An indexed state does not guarantee a standard search result, an AI-feature supporting link, or a citation in an answer.
What one site audit showed
On November 11, 2026, ordinary public GET requests to the resources page and the sitemap audit article returned HTTP 200. Both responses contained self-referencing canonical tags and no noindex robots meta tag. The requests used a general audit user agent, not Googlebot or an AI-provider crawler.
For the sitemap audit article, 15 selected visible paragraphs longer than 60 characters also appeared in normalized text from a fresh, unauthenticated HTML response. The comparison excluded the comment form. It shows that the server returned those selected passages at the time of the check. It does not show that a search engine or answer provider fetched, indexed, ranked, or cited the page.
The same check found that /sitemap_index.xml returned XML with a urlset root element and 827 immediate URL elements. The filename did not determine the XML structure; the returned document did. That observation applies only to that site on that date. It does not support a general conclusion about Yoast or other sites.
When browser content and response HTML differ, compare matching sections before drawing conclusions. Exclude login-only areas, comment forms, and other interactive elements when they do not represent public page content. Then check whether script rendering explains the difference. A result for one URL does not describe every URL on the site.
Run the access check
Check the exact URL a visitor or crawler requests, including the final destination after redirects. Record the status code, redirect chain, canonical URL, and any noindex instruction in the HTML or HTTP headers. Confirm whether the URL appears in a submitted sitemap. Then compare important page text in the server response with text shown in a browser.
Fix a specific access or indexing-instruction problem before interpreting visibility data. When these checks pass, they provide evidence about the page response and site configuration. Provider tools or logs establish provider activity. Search-result and answer observations establish visibility.
Counterargument: a 200 response still has diagnostic value
A successful public response remains useful evidence. It can expose broken redirects, accidental noindex instructions, conflicting canonical tags, missing server-rendered text, or a sitemap that lists the wrong URL. Its limit is clear: the response describes the audited request, while crawler access, indexing, and answer selection require separate proof.
Measure organic search
Use Google Search Console’s Performance report to measure how often pages from your site appear in Google Search and how often searchers click them. The report shows impressions, clicks, queries, pages, and available dimensions such as country, device, and search appearance. Compare equivalent date ranges and use the same search type in both periods. Google defines the report’s metrics, filters, and limits in its Performance report documentation.
Google’s AI features documentation, updated December 10, 2025, states that traffic from AI Overviews and AI Mode is included in overall Search traffic. Search Console reports that traffic under the Web search type. The documentation does not describe a filter that separates AI Overviews or AI Mode from other Web results. Treat a Search Console Web total as a measure of Web search performance, not as a measure of AI-feature traffic.
Use website analytics to measure what visitors do after they arrive. Review sessions, landing pages, available traffic-source dimensions, and configured events such as form submissions, purchases, or account sign-ups. Google Analytics documents traffic acquisition reports and event measurement. Analytics records activity on your site; it does not record how often Google showed a page or mentioned it within an answer interface.
Keep indexing, search performance, and on-site outcomes separate. An indexed page is a page Google has added to its search index. An impression records that the page appeared in a qualifying Google Search result. A click records a visit from that result. A conversion records a defined on-site action, such as a purchase or lead submission.
Indexing reports show Google’s status for a page. The Performance report shows activity in Google Search. Website analytics shows activity after a visitor arrives. A rise or fall in one measure does not, by itself, explain movement in the others.
Track AI answers consistently
Treat AI answers as repeated observations, not as a complete record of what every user sees. Keep branded and unbranded questions separate. Branded questions name the company or product and show how an answer product describes it. Unbranded questions test whether the company or product appears during broader customer discovery.
Build a core panel from real customer questions and buying decisions. Group questions by intent, such as learning about a category, comparing products, checking a claim, or deciding what to buy. Keep the wording, language, intent, and topic mix stable when comparing periods. Put new questions in a separately labeled exploratory panel, and record when a question enters or leaves the core panel.
Use a fixed observation record
Record the panel ID and whether the group was branded or unbranded, plus the question ID and exact wording. Also record the answer product and interface or mode, the locale and language, date and time, and run number. Preserve the raw answer, displayed source labels and URLs, and reviewer notes so another reviewer can understand what was tested and reproduce the observation.
Record the brands and products named in the answer, claims about the business, recommendation status and context, citation status, and any assessment of accuracy or source support. Preserve the raw answer before coding it. Keep coding definitions and document later corrections so reviewers can see how a mention, recommendation, or supported claim was counted.
For example, “Brand mentioned,” “recommended for small businesses,” and “linked to the company’s pricing page” are separate observations. A displayed link shows that the answer product presented a source. It does not show that the source supports every claim in the answer. Check important claims against the cited page and record whether the page supports, contradicts, or does not address each claim.
Save a screenshot or export when permitted and practical, especially when the interface displays answer text, citations, or links that a later export might omit. Record visible conditions that shape the result, including the model or mode when known, browsing or search settings, and relevant account or interface details.
Keep consumer-interface observations separate from API results. An API response describes one configured API request at a particular time. It does not describe every consumer experience that uses a similar model name.
Repeat important questions under the same recorded conditions. Retain every completed result rather than selecting the most favorable answer. Record failed, unavailable, or missing runs separately.
Asking one question five times gives one distinct question with five runs; asking five questions once each gives five distinct questions with one run apiece. These measures answer different questions. For a rate across runs, divide completed runs with the outcome by all completed runs. For a rate across questions, divide distinct questions with the outcome by all tested questions. State which rate you report and show its denominator.
Keep a comparable panel
Use the same core questions for period-to-period comparisons. When the panel changes, report the old and new panels separately, or recalculate both periods using only questions included in both. Keep a change log with the date, reason, and question ID for every addition or retirement.
Show counts beside percentages. A small panel is sensitive to individual results: one changed answer can materially shift the percentage.
Illustrative arithmetic only: If a brand appears in 3 of 10 tested questions, its observed share is 30%. If one question is added and the brand appears in 4 of 11 questions, the observed share becomes about 36%. That increase does not establish improved visibility because the question mix changed.
Report intent groups separately when they matter. Strong results for category-learning questions can conceal weak results in product comparisons or purchase decisions. Keep the exploratory panel visible for emerging customer concerns, but do not silently blend it into the core trend line.
A fixed panel has clear limits
A fixed panel measures answers for the tested products, modes, locales, languages, customer intents, and dates. It does not measure every query or every buyer experience. State those boundaries beside reported rates and trends.
The same limit applies when an answer product does not display a result type for a tested question. Google’s documentation on AI features states that AI Overviews do not appear for every query. The absence of an Overview in a sample does not show that the measurement failed.
Counterargument: Why not track every new question?
A broad, changing question set finds emerging customer concerns faster than a fixed panel, making it useful for exploration. Trend comparisons require a stable question mix because changing questions changes the measured rate. Use an exploratory panel for discovery and a stable core panel for comparison.
Code AI mentions, citations, and recommendations
A brand name in an AI answer, a buyer-relevant recommendation, and a displayed source link are separate outcomes. Record each separately. A citation shows that the interface displayed or named a source. It does not establish that the answer represented the source accurately, recommended the business, or sent a reader to the business’s website.
Define what counts
Count a mention when the answer names the business, brand, or product. Save the surrounding sentence. If the wording evaluates the offer, label the evaluation positive, negative, neutral, or unclear. A criticism and a passing reference both count as mentions. Neither counts as a recommendation on its own.
Count a buyer-relevant recommendation when the answer presents the business or product as a suitable choice for the needs stated in the question. Record the use case, reason, qualifications, and alternatives named. A business included in a provider list does not qualify unless the answer connects it to a buyer constraint, such as budget, location, required feature, or eligibility.
Count a citation when the answer displays a link or identifies a source. Record whether the source is the business’s own page or a third-party page that mentions the business. A third-party article counts as a citation to that article, but does not show that the answer cited the business directly or endorsed it.
Set the counting unit before reviewing answers. For example, count a brand once in each question-and-answer record even if its name appears several times. Count individual links only when link volume is the intended measure. State the rule and denominator so repeated names or links do not inflate results.
| Outcome | Count it when | Do not count it when |
|---|---|---|
| Mention | The answer names the business, brand, or product. | The name does not appear. |
| Recommendation | The answer presents it as a fit for the buyer’s stated needs. | The answer only lists it among alternatives or mentions it in passing. |
| Citation | The answer displays a link or identifies a source. | No source appears. |
Check cited claims one by one
A displayed citation does not prove that its page supports the answer. For each important claim tied to a source, open the cited page and compare its wording with the claim. Save the claim, URL, relevant passage, and review decision.
Use four support labels:
- Direct: The page clearly supports the claim as stated.
- Partial: The page supports part of the claim but omits or limits a material detail.
- Absent: The page does not support the claim.
- Unclear or not assessable: The page is unavailable, ambiguous, or lacks enough information for a sound decision.
Flag stale information, ambiguous wording, and contradictions. Check high-stakes claims, including price, eligibility, availability, safety, and legal status, against the cited page and, where needed, current business records.
Keep source support separate from answer accuracy. An answer can misrepresent a relevant source. It can also make an accurate claim without a visible citation. Record both findings; a citation alone does not establish that the claim is true.
Apply the same coding rules to every run
Define what qualifies, what does not, and how reviewers handle borderline cases before coding a batch. Use unclear or not assessable when the answer lacks enough evidence. Forced judgments create false precision.
Code sentiment only when the answer clearly evaluates the brand or its offer. Preserve the wording behind each positive or negative label. A statement that an offer costs more does not establish negative sentiment when the answer presents that price as appropriate for a premium buyer. Review automated sentiment labels against the saved text because automated classification can miss that context.
When more than one reviewer codes answers, have each reviewer code the same sample independently. Compare decisions, resolve disagreements, revise the definitions where needed, and recode affected records consistently.
Report counts and shares separately for mentions, recommendations, and citations. Break results out by product or mode, question group, and reporting period when the sample supports the comparison. Keep branded prompts separate from unbranded prompts because they measure different forms of visibility.
For example, “3 of 20 answer runs recommended the business” is interpretable when the question panel, product or mode, and period are stated. “Three recommendations” is not. If the prompt panel changes, report the change beside the results because a different prompt mix changes counts and percentages.
Keep answer evidence separate from website visits
A link shown in an answer and a referral visit are different observations. Record citations from saved answer runs. Record visits through web analytics, referral data, and landing-page records.
Google Analytics traffic-source dimensions, including source and medium, describe the traffic information Analytics receives, as documented in Google’s traffic-source reference. Those dimensions do not identify the specific AI answer that prompted a visit. Some visits arrive without a usable referrer and are classified as direct traffic or remain unclear.
Ask new leads, “How did you first hear about us?” and retain the response as self-reported discovery. The answer adds context, but memory and question wording affect it. Do not use it to overwrite analytics records or claim that one answer produced a sale.
Keep consumer products and APIs in separate reporting rows. A consumer product includes its own interface and user settings. An API result reflects the request, model selection, and configuration used for that call. They are different observations.
Use a small scorecard
A monthly scorecard should show each measure, its denominator, and its limits.
| Measure | Report |
|---|---|
| Answer observations | Mentions, buyer-relevant recommendations, and citations, each with counts and denominators |
| Source support | Claims checked and the share with direct, partial, absent, or unclear support |
| Website outcomes | Referral sessions, qualified leads, and revenue, with attribution limits stated |
| Question coverage | Branded and unbranded prompts, locale, product or mode, and date range |
Keep a change log for page edits, pricing changes, promotions, and product launches. Also log prompt revisions, locale changes, product-mode changes, and coding-rule updates. The log helps interpret a shift in results. It does not prove that a recorded change caused a business outcome.
A fixed question panel gives a repeatable sample of selected buyer questions. It does not represent every question a buyer might ask. Check important findings against customer conversations, cited pages, and business records before acting on them.
Counterargument: Why not use one visibility score?
A single score is easy to display, but it hides the difference between being named, recommended, and cited. A business may receive many mentions because it appears in comparison lists while receiving few fit-based recommendations. Another may receive citations without being named. Combine measures only when the formula, weights, and purpose are disclosed. Otherwise, report each outcome separately.
Connect visits to business results
Website analytics records activity after someone reaches your site. It can report sessions and landing pages, along with available referral information. Configured events can include form submissions or purchases. The records depend on the tracking setup and consent choices. Browser behavior and the data sent to the analytics product also shape them. They do not count people who saw a brand in an AI answer and never visited the site.
A referral session means the analytics record received information about a referring source. It does not show how many people saw an answer, count unlinked brand mentions, or capture every visit influenced by an AI product. Browser and site settings also limit the referral information sent between sites. The Referrer Policy standard describes how websites can limit that information. Keep direct, unassigned, and unknown traffic in those categories. Do not reclassify it as AI traffic because an answer mentioned the business.
Keep recorded visits, self-reported discovery, and CRM attribution in separate fields. In the CRM, define qualified lead, closed sale, and recognized revenue according to the business’s own rules. Apply those definitions consistently across reporting periods.
Ask new leads or customers a short, optional question such as, “How did you hear about us?” Keep response choices stable and include a free-text option. Report completed responses and skipped questions. Self-reported answers can reveal discovery that analytics did not record, but recall, question wording, and nonresponse limit what those answers establish.
Revenue attribution also requires a stated rule. For example, a business might associate a purchase with an AI referral when analytics records that referral before the purchase within a defined 30-day window. Report the rule and window beside the result. The result is revenue attributed under that rule; it does not prove that an AI answer caused the purchase. A buyer may see an answer, search for the brand later, and arrive through another channel.
Do not use referral sessions or direct traffic as measures of brand awareness or loyalty. Measure awareness with a consistent survey. Measure loyalty with a defined business outcome, such as repeat purchases or renewal rates.
Follow the conversion path
Link an analytics form-submission or purchase event to a lead or order ID when the consent model and data-handling rules permit it. Then connect that ID to the CRM’s qualification status and, for completed sales, to revenue in CRM or commerce records. Document how duplicate form submissions, canceled orders, refunds, and returns are handled so revenue follows the same rules in every reporting period.
Keep first-touch and last-touch fields only when their definitions are clear. First touch records the earliest known source under the business’s tracking rules; last touch records the latest eligible source before conversion. Neither field captures every influence on a buyer’s decision.
A compact reconciliation view places four measures side by side:
| Measure | What it records |
|---|---|
| Answer visibility | Mentions or recommendations in a fixed set of sampled questions |
| Recorded visits | Analytics sessions and landing pages by available referral information |
| Business results | Qualified leads, orders, and revenue linked under the stated attribution rule |
| Self-reported discovery | Survey responses, including response and nonresponse counts |
The totals will not match because each measure describes a different stage and relies on different records. A sampled answer observation does not measure audience reach. A referral session does not capture every AI-influenced visit. A self-reported answer reflects what a respondent remembers and chooses to share. Treat the gaps as unknowns rather than forcing the measures into one total.
Read each number for what it measures
Analytics reports only what the tracking setup receives. For example, Google Analytics describes its Traffic acquisition report as a session-acquisition report. Configured events record actions that the site sends to Google Analytics, such as purchases or form submissions. Verify that relevant events fire correctly and that collection practices meet the business’s consent and data-handling requirements.
Keep the sampled-question panel and reporting definitions stable between periods. If the questions, answer-checking method, attribution window, or lead definition changes, identify the change beside the results. A rise in recorded referral sessions warrants a closer look at landing pages, lead quality, and downstream purchases. It does not establish that greater answer visibility caused higher sales.
Why not use one attribution number?
A single last-touch figure is easy to compare, but it answers a narrow question: which eligible recorded source came last under the stated rule? The figure remains useful for a specific budget or channel decision. It becomes misleading when presented as a complete account of how customers discovered the business.
For example, a buyer may first encounter a business in an AI answer, later search for its name, and then purchase after a visit that analytics assigns to organic search, paid search, email, or direct traffic. Keep the last-touch figure, self-reported discovery, and known earlier touchpoints separate rather than treating one field as the full customer journey.
Does an AI referral prove that an answer produced a sale? No. It records a visit with available referral information. A purchase linked to that visit is revenue attributed under the business’s stated rule, not proof of causation.
Should direct traffic count as AI discovery? No. Direct traffic has no recorded eligible source under the analytics setup. Record it as unknown unless another reliable record identifies the source.
Use answer checks to describe visibility under sampled conditions, analytics to describe recorded visits, and CRM or commerce records to define qualified leads and revenue. These measures support better decisions when each retains its own definition and limits.
Build an SEO and GEO scorecard
For SEO and GEO reporting, place access checks, search performance, sampled answer outcomes, and business outcomes in one report while keeping the evidence for each measure separate. Each row should define the measure and show its current value and denominator. It should also identify the comparison period and source, state coverage, and name the main limitation.
| Measure | Current value and denominator | Comparison period | Source and coverage | Caveat |
|---|---|---|---|---|
| Access checks | [x of n URLs checked] |
[dates] |
HTTP responses, page directives, and canonical tags for [URL set] |
A successful request does not show that a search provider fetched, indexed, cited, or ranked the page. |
| Indexing checks | [x of n URLs indexed] |
[dates] |
Search Console URL Inspection or indexing reports for [URL set] |
This is Google-specific evidence. It does not describe every search engine. |
| Search clicks | [count] |
[date range] versus [comparison range] |
Google Search Console, Web search type | Clicks record Google Search activity. They do not identify the AI feature, if any, that led to the click. |
| Search impressions | [count] |
[date range] versus [comparison range] |
Google Search Console, Web search type | Google includes AI Overview and AI Mode traffic in overall Web reporting without separate filters for either feature, according to Google’s AI features documentation. |
| Answer-panel coverage | [x of n questions; y%]; [completed runs of planned runs] |
[dates] |
Saved answers, with product or mode, panel version, and locale | State whether the rate uses distinct questions or repeated runs. Record unavailable responses and failed runs. |
| Brand mentions | [x of n answers; y%] |
[dates] |
Coded answer records, reported separately for branded and unbranded prompts | A mention does not show that the answer recommends the business. |
| Buyer-relevant recommendations | [x of n answers; y%] |
[dates] |
Coded answers grouped by buyer intent | Define the buyer need and the conditions that qualify as a recommendation. Report unclear cases separately. |
| Direct and third-party citations | [direct x of n answers]; [third-party x of n answers] |
[dates] |
Displayed links and source labels in saved answers | A displayed citation records a source shown with the answer. It does not prove that the source determined the answer’s wording. |
| Source support and answer accuracy | [supported claims of claims reviewed]; [accurate claims of claims assessed] |
[dates] |
Claim-level review against cited passages and reliable evidence | Define “supported” and “accurate.” Keep claims that cannot be assessed separate rather than forcing a judgment. |
| Recorded referral sessions | [sessions] |
[date range] versus [comparison range] |
Website analytics, referral reports, and landing-page reports | Privacy controls and missing referrer data leave some visits unattributed. Do not label unknown traffic as AI referrals. |
| Qualified leads | [count] |
[date range] versus [comparison range] |
CRM or lead records using a stated qualification rule | Keep the qualification rule and reporting window consistent. |
| Conversions and revenue | [count]; [$ amount] |
[date range] versus [comparison range] |
Commerce or CRM records, with the attribution rule stated | Revenue attributed to a recorded referral does not prove that an answer caused the sale. |
Interpret the numbers beside the table
Add a short plain-language note for each reporting period. State what changed, which observations support that reading, what remains uncertain, and which decision follows.
For example, with a stable question panel, a higher recommendation rate shows that more sampled answers met the stated recommendation rule. The rate alone does not explain why it changed, whether buyers saw the same answers, or whether the change produced revenue. Review the underlying answers, repeat the affected prompts, and check whether cited pages contain the needed product information.
Keep a change log
Record each relevant edit with its date, affected URL or external source, exact change, reason, owner or role, publication status, and related question-panel IDs. Record technical changes separately from content changes. Technical changes include redirects, canonical tags, and robots rules. Content changes include revised page copy, corrected product specifications, and third-party listing corrections.
Also record observation dates before and after the edit, products or modes tested, locale, panel version, and concurrent changes such as promotions, pricing updates, campaigns, or product availability. The log connects later observations to earlier work and helps select the next check.
A later visibility change establishes timing, not cause. Use the log to form a hypothesis, then repeat the relevant checks and look for concurrent changes that offer another explanation.
A composite score can hide the decision
A single visibility score seems convenient when a report contains many measures. It collapses unlike evidence: an HTTP status, a Search Console impression, a cited answer, a referral session, and revenue carry different meanings and levels of certainty.
Report the separate measures. If an internal composite helps prioritize work, disclose every input, weight, and limit, and retain the underlying rows. Do not compare the result with a universal benchmark unless it uses a matching question panel and product mix, covers the same locale, and follows the same measurement method.
Common questions
Should results from different answer products or modes be combined?
Report them separately first. Products and modes can use different interfaces, source displays, and answer behavior. A combined total can conceal a meaningful change in one product.
Does a later visibility change prove that an edit worked?
No. A change after an edit shows sequence, not cause. Use the change log to form a hypothesis, repeat the same panel, and check for concurrent changes.
A useful scorecard marks unavailable data as unavailable rather than treating it as zero. It preserves the denominator behind every rate, shows what the evidence supports, and points to the next decision.
Interpret SEO and GEO results
Use the panel as a diagnostic
A small, repeated panel of buyer questions can reveal problems worth fixing. An answer may describe a product inaccurately, omit information buyers need, or cite a page that does not support its claim. Repeated tests may also show that a business remains absent from a defined set of questions.
These findings describe the panel and the conditions under which you tested it. They do not show how many people saw the answers, how often a brand appeared across all searches, or total market exposure. Those claims require a sampling design that supports estimates beyond the tested questions.
Keep observations that conflict with earlier results. Report the number of questions and runs behind each finding. In a small panel, one changed answer can shift a percentage sharply. One missing mention is evidence from one observation, not proof that the business is invisible.
Keep Google features separate
Google says AI Overviews appear when its systems determine they add value to Search, so they do not appear for every query. An absent AI Overview does not show that tracking failed.
Google also says that AI Overviews and AI Mode may use different models and techniques. Its AI features can identify multiple supporting pages. A citation pattern in an AI Overview therefore does not establish what AI Mode will cite, and an AI Mode result does not establish the pattern for AI Overviews. These statements describe Google Search features, not every answer engine.
Record the product, feature, mode, locale, question wording, date, and run for each observation. Separate results when any of those conditions change. If the question set changes between periods, counts and percentages may change because the panel changed rather than because answer behavior changed.
Treat access checks as limited evidence
An access check answers a narrower question than an answer observation. If a public request to a resource page returns HTTP 200, the server responded successfully to that request at that time.
That result does not show that a provider crawler fetched the page, that a search engine indexed it, or that an AI feature used or cited it. A request made with a general audit user agent describes only the response to that user agent. Record the URL, date, request method, and user agent, then limit the claim to what the check tested.
Do not claim cause from sequence
A visibility change after a content edit establishes timing, not causation. Search demand, competing pages, product changes, answer-system updates, and other marketing activity may have changed during the same period. Record those conditions beside the edit. Describe the result as an association unless a suitable comparison supports a stronger conclusion.
Google has said that clicks from Search result pages with AI Overviews are “higher quality,” which it describes as users being more likely to spend more time on the destination site. Google does not state the sample or method behind that observation in the announcement. Treat it as Google’s observation about its Search traffic, not as a benchmark for a particular business or site.
For a decision with material financial, legal, or operational consequences, have a relevant professional validate the underlying analytics and customer-record data. That professional should work in analytics, privacy, finance, or law.
When noisy results still help
Variation does not make a small panel useless. A panel supports a decision when it uses a defined set of buyer questions under consistent conditions and produces a repeated finding: an inaccurate product description, a missing buying detail, weak source support, or a persistent absence from that panel.
Those findings justify checking the relevant page, product information, and supporting sources. They do not support claims about all customers or all searches. Their value lies in identifying a specific issue that recurs under documented conditions, then verifying whether the business can correct it.
Questions readers ask
Can Search Console isolate AI traffic?
As of Google’s AI features documentation, updated December 10, 2025, clicks from AI Overviews and AI Mode appear in overall Search traffic under the Web search type. Google documents no separate Search Console report or filter for isolating traffic from either feature.
Check the current Search Console interface before describing available reporting, since Google can change the interface. A Web total does not reveal how much traffic came from AI Overviews or AI Mode.
Does a citation mean a recommendation?
No. A citation means the answer displayed a link to a source. A recommendation presents a business, product, or service as a suitable choice for the buyer’s stated need.
Record citations and recommendations separately. Also check whether the cited page supports the claim in the answer. An AI answer may cite a page for a factual detail without presenting that business as a purchase option.
How many prompts should I track?
Track a manageable set of questions that covers the customer needs, search intents, and buying decisions that matter to the business. Keep a stable core set so results remain comparable over time, and repeat the questions that matter most.
Report the number of prompts, the number of runs, and the panel’s coverage. Results describe the selected prompt set; they do not represent every possible query or every buyer.
Can analytics show all AI discovery?
No. Analytics records visits and configured events after someone reaches the site with usable tracking information. It cannot record an unclicked mention in an AI answer. Missing, stripped, or limited referral data can also leave the source of a visit unknown.
Ask leads or customers the same question about how they first heard about the business. Keep self-reported answers separate from observed referral traffic. The first records what the visitor says; the second records what tracking data identifies.
Key Takeaways
- Separate the evidence layers: Track page access and indexing, organic search performance, AI-answer observations, and business outcomes as distinct measures. An HTTP response, indexed page, or AI mention does not prove a sale.
- Keep answer tests comparable: Separate branded from unbranded questions, retain the same core panel, and record each product or mode, locale, date, raw answer, sources, and repeated runs. Report counts with percentages and denominators.
- Code answer outcomes separately: A mention, buyer-relevant recommendation, displayed citation, and well-supported claim are different findings. Check cited sources against the claims they are meant to support.
- Connect visits to outcomes with clear limits: Use analytics for recorded visits and configured events, CRM or commerce records for qualified leads and revenue, and optional self-reported discovery to identify missing attribution. Log changes and concurrent events; timing alone does not establish causation.
FAQ
Can Google Search Console isolate AI Overview or AI Mode traffic?
As of Google’s documentation updated December 10, 2025, clicks and impressions from AI Overviews and AI Mode are included in overall Search traffic under the Web search type. The documentation does not describe a separate filter that isolates those features, so do not label a Search Console total as AI-feature traffic.
Does a citation mean my business was recommended?
No. A citation records a displayed source link, while a recommendation means the answer presents your business as a fitting choice for the question’s needs. Record mentions, recommendations, and citations separately, then check whether each cited page supports the claim made.
How many prompts should I track?
Choose questions that cover important customer needs, buying decisions, and topics, then report how many questions and runs you tested. No universal sample size represents every customer or query. Repeat important questions and describe which products, modes, locales, and intents your panel covers.
How do I keep answer tracking comparable over time?
Keep a fixed core panel: use the same wording, product or interface, mode, and locale. Record the date and each repeated run. Save the raw answer and displayed sources before coding mentions, recommendations, accuracy, or source support. Report counts and percentages with denominators.
What if my question panel changes?
Show results from the unchanged core panel separately from results for new or retired questions, or recalculate both periods using the same questions. A changed mix can move counts and percentages even if the answers themselves are no better.
Can a basic HTTP check prove that a search engine or AI product can use a page?
No. A successful response, visible page text, and the absence of a noindex directive describe what an ordinary request received at that time. They do not prove that a provider crawler fetched or indexed the page, or that an answer product cited it. For example, a November 11, 2026 check of two pages on vincentschmalbach.com returned HTTP 200 and found self-referencing canonicals, but it used a general audit user agent rather than a provider crawler.
Can analytics show all AI discovery?
No. Analytics can record visits and configured events when people reach the site with tracking information, but it cannot count unclicked mentions. Missing or limited referrer data can also obscure where a visit began. Add an optional, consistently worded “How did you hear about us?” question, and report those answers separately from recorded referrals and CRM-attributed leads or revenue.
Should I treat an API answer and a consumer app answer as the same result?
No. Record the product and interface or mode tested, along with any known model, browsing, search, and locale settings. An API result describes that configured call; it does not establish what users see in a consumer app with a different interface or configuration.
What I'm buildingDelegate tasks. Get software.
Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.
Take a look at vroni.com