Share of answer is the operating metric for this practice.

How we count it
Request an audit

Insight · 12 min

Measuring AI search visibility

To measure AI search visibility, run a fixed, reviewed query set on a fixed panel of engines, repeat each query because answers change between runs, and log the answer, the sources, and the date every time. Score mentions, citations, links, and accuracy as separate numbers. Then read first-party data from Search Console, Bing Webmaster Tools, analytics, and server logs beside the panel, because neither source is complete alone.

In short

  • The panel is the measurement. First-party reports show impressions, citations, and clicks, but not what the answer said.
  • Answers are non-deterministic. One run per query is an anecdote. Repeat runs and report a range.
  • Log every run the same way: query, engine, settings, answer text, sources, position, date, screenshot.
  • Keep mention, citation, link, and accuracy apart. Blending them hides the fix.
  • Treat movement as association. Record the baseline, the observation window, and every concurrent change.

What does “AI search visibility” mean in practice?

AI search visibility is how often, and how accurately, your brand and your URLs appear in generated answers for the queries your buyers ask. The headline metric is share of answer: the share of a defined query universe where you are the cited source. This piece is the method that produces that number.

Two kinds of evidence feed it. Observed evidence comes from running queries and recording what the engines return. First-party evidence comes from the reports engines and your own systems give you. The first tells you what was said. The second tells you how often you were shown or clicked. You need both.

How do you build the query set?

Build the query set from what buyers ask, reviewed by someone who sells the thing, and freeze it before the baseline. A set that changes every month cannot show a trend.

  1. Start from sales calls, support tickets, site search, and Search Console queries. Keyword tools are an input, not the list.
  2. Tag each query with one intent: commercial, definitional, procedural, or comparative.
  3. Write each query the way a buyer would type or say it. Include longer, conversational forms; generated answers are often triggered by them.
  4. Remove vanity prompts. A query that contains your brand name measures recall of your brand, not visibility in the category. Keep a small branded set, labelled and reported apart.
  5. Size it so it can be run on cadence. A reviewed set of 100 to 300 queries is usually more useful than an unread export of thousands.
  6. Version the set. When a query is added or removed, record the date and the reason.

Which engines belong on the panel?

The panel should contain the products your buyers actually open, chosen from evidence, not from the trade press. For most B2B and consumer categories that means some of Google AI Overviews, Google AI Mode, ChatGPT search, Perplexity, Gemini, Microsoft Copilot, and Claude. Each retrieves from a different search system; see where each engine gets its sources.

Check referral data and ask customers before you choose. An engine that sends no traffic and that no buyer mentions can come off. Record the engine, the product surface, and the model or mode where the product shows it, because the same brand name can cover several systems.

How should the runs be set up?

Control everything you can, and write down everything you cannot. The goal is that a change in the log reflects a change in the engine or the web, not a change in how you asked.

  1. Account state. Run logged out by default, in a clean browser profile. Personal history and memory features can shape answers. If buyers mostly use a logged-in product, add a logged-in run with a fresh account that has no history, and label it.
  2. Location and language. Fix the country, region, and interface language. Use the same network location each run. Local queries need the buyer’s location, not the office’s.
  3. Device. Pick desktop or mobile per engine and keep it. Layout changes what is visible.
  4. Phrasing. Paste the exact stored query. Do not ask follow-ups unless the follow-up is part of the stored query. In Google AI Mode, a follow-up is counted as a new query in Search Console, so treat it as one in the panel too.
  5. Repeats. Run each query several times per session, and on more than one day. Three to five runs per query per cycle is a practical floor for a small set. Report the share of runs where you appear, not a yes or no.

Repeats are not optional. Research on Perplexity, SearchGPT, and Gemini found that citation results vary substantially across repeated samples, and that many apparent differences between domains fall within the noise of the measurement. The authors conclude that single-run visibility metrics give “a misleadingly precise picture.” A separate 2026 paper makes the same argument in its title: measure visibility as a distribution, not a single-point outcome. If two engines or two months differ by less than the run-to-run spread, report no change. How many queries and runs you need for a given margin, and how to test a change between periods, is worked through in share of answer.

What should you log for each run?

Log one row per query, per engine, per run, with enough detail that someone else could check it later. The platform keeps these in one record; a spreadsheet works if the columns are fixed.

FieldWhat to record
Run IDUnique ID, query set version
Date and timeWith time zone
QueryExact stored text, intent tag
Engine and modeProduct, surface, model or mode if shown
SettingsLogged in or out, location, language, device
Answer presentFor Google: whether an AI Overview appeared at all
Answer textFull text, copied, not summarized
Brand mentionedYes or no, and which brand names
Cited sourcesEvery source URL, in display order
Your positionOrder of your first citation among sources, if cited
LinkWhether your citation is a working, clickable URL
AccuracyCorrect, outdated, wrong, or unsupported, with the sentence
CompetitorsBrands mentioned and domains cited
ScreenshotFile reference, full answer and source panel

Copy the text as well as the screenshot. Screenshots prove what was shown; text can be searched and compared across runs.

Score four things separately, per query and per engine, as the share of runs in which each occurred. The distinction is set out in what gets cited; here is how to apply it.

  • Mention: the brand appears in the answer text.
  • Citation: a page you control is attached as a source.
  • Link: that citation is a working URL a person can click.
  • Accuracy: every sentence about you matches the current offer.

A query where you appear in four of five runs scores 0.8 for that measure on that engine. Roll up to the query set by intent, never across measures. Score accuracy against a written reference: current prices, plans, features, and claims, with the date of each. Check whether the cited source supports the sentence. A 2023 audit of four generative search engines found that on average only 51.5% of generated sentences were fully supported by their citations, and only 74.5% of citations supported the sentence they were attached to. A citation is not proof that the answer is right.

What does Google Search Console show for AI Overviews and AI Mode?

As of November 2026, Search Console counts AI Overviews and AI Mode inside the Performance report’s “Web” search type, and adds a separate Generative AI performance report that shows impressions only.

In the standard Performance report, Google’s documentation says that clicking an external link in an AI Overview or AI Mode counts as a click, that standard impression rules apply, and that all links in an AI Overview share the overview’s single position. Position in AI Mode follows the same method as a results page. A follow-up question in AI Mode is treated as a new query. Search Labs experiments are excluded. The standard report does not separate AI feature traffic from other web results.

The Generative AI performance report, announced on June 3, 2026 and rolled out to all websites as of August 31, 2026, shows how many times links to your site were shown in generative AI features on Google Search. Its help page lists pages, countries, devices, and dates as dimensions. It does not list queries, clicks, or position, and it reports AI Overviews and AI Mode together. Google also offers a Search generative AI control that removes a site from these features; excluded sites receive no impressions or traffic from them.

Use it for trend and page coverage. It cannot tell you which query triggered the impression or what the answer said. Rankings held, clicks fell shows how to work out which queries show an overview.

How do ChatGPT and Perplexity referrals appear in analytics?

ChatGPT referrals usually arrive tagged with utm_source=chatgpt.com; OpenAI’s publisher FAQ says ChatGPT adds that parameter to referral URLs automatically. Perplexity, Gemini, Copilot, and Claude show up as referral domains when a referrer is passed, and as direct traffic when it is not. As of November 2026, GA4 also has a default AI Assistant channel, added in May 2026, which does not publish its full source list and leaves AI Overviews and AI Mode in Organic Search.

Referral data measures clicks. Most answers produce no click, so it is the smallest and least complete of the numbers on the panel. The channel setup, the “how did you hear about us” field, CRM capture, and the pipeline report are covered in AI search attribution.

Does Bing Webmaster Tools report Copilot citations?

Yes. Bing Webmaster Tools has an AI Performance report, in public preview since February 10, 2026, that counts how often your pages are cited as sources in Microsoft Copilot, AI-generated summaries in Bing, and “select partner integrations.” Microsoft does not name the partners. The launch post does not mention ChatGPT, so do not read the report as ChatGPT data.

The launch post lists five views:

ViewWhat Microsoft says it showsWhat it does not show
Total citationsHow many times your content was shown as a source in AI answers in the selected periodWhere in the answer the citation sat
Average cited pagesThe daily average of unique pages from your site cited as sourcesRanking, authority, or a page’s role in the answer
Grounding queriesThe key phrases the AI used when it retrieved content it then citedThe user’s prompt. Microsoft says this view “represents a sample of overall citation activity”
Page-level citation activityCitation counts per URL over the date rangePage importance, ranking, or placement, in Microsoft’s words
Visibility trendsA timeline of citation activity across supported AI experiencesWhy a trend moved

A June 16, 2026 update, also in preview, added four things. Intents classify grounding queries into categories such as informational, commercial, and navigational. Topics group related grounding queries into clusters. Citation share is the percentage of citations attributed to your site out of all citations shown across all sites for a grounding query. Compare overlays an earlier period on the current one. Microsoft calls citation share “an observational metric – not a ranking system or a competitive scoreboard,” and says it does not expose competitor domains or represent traffic share.

What the report lacks matters as much as what it has. Neither post mentions clicks, click-through rate, or impressions for AI answers, the answer text, or the position of your citation inside the answer, and neither describes an export or API. Grounding queries are a sample, and they are the phrases the system searched with, not what the user typed. As of November 2026 the report is still labelled preview.

Use it the way you use the Search Console generative AI report: for trend and page coverage on one engine family. It is the only first-party citation count on a typical panel that we could verify as of November 2026. It counts citations, not clicks, and it cannot tell you whether the cited sentence was accurate. Map its grounding queries against your query set; a grounding query you do not track is a candidate to add, with a version note.

What can server logs show?

Server logs show which AI agents fetched which URLs, and when. They show access, not citation.

Split AI user agents by purpose. OpenAI documents three: OAI-SearchBot for ChatGPT search, ChatGPT-User for fetches a user’s action triggers, and GPTBot for model training. Perplexity documents PerplexityBot for search results and Perplexity-User for user-initiated fetches. Both companies publish IP ranges, so verify that a request claiming to be a bot comes from the published addresses. Google names Googlebot as the crawl control for AI features in Search, so logs cannot isolate AI Overview fetches.

A rising count of user-triggered fetches on a page suggests that page is being retrieved for answers. A page that search agents never fetch is unlikely to be cited by that engine. If fetches are absent, check access policy first; see AI crawlers in robots.txt.

What do third-party AI visibility trackers measure?

Third-party trackers automate the panel: they send prompts to engines on a schedule and parse mentions and citations. They save time. They also carry limits that should be stated in any report built on them.

  • The prompts are synthetic. They are written by you or the vendor, not drawn from real user sessions, and volume estimates for them are modelled.
  • Coverage is sampled. Run counts, locations, and account states vary by vendor and are not always disclosed.
  • Many run through APIs rather than the consumer product. API answers and the logged-out product can differ.
  • Parsing can misread brand names, especially generic ones.

Ask any tool for its run count per prompt, its location and account settings, and whether it captures the consumer interface. The same questions apply to agencies; see how to choose a GEO agency. Spot-check a sample against manual runs each quarter.

What cadence and report format work?

Run a small volatile set weekly and the full set monthly, and report quarterly against the baseline. Volatile means commercial and comparative queries, or queries where a competitor is cited.

The monthly report fits on one table per engine, by intent:

IntentQueriesMentionCitationLinkAccurateChange vs baselineRun spreadConcurrent changes
Commercial
Definitional
Procedural
Comparative

Beside it, put the first-party numbers for the same window: Search Console web clicks and generative AI impressions, Bing citations, AI referral sessions, and AI agent fetches of the target pages. Close with the list of commercial queries where a competitor is cited and you are not, and the reason for each.

What are the common measurement mistakes?

Most bad readouts come from a handful of repeatable errors.

  • Vanity prompts. Queries that name your brand inflate every number and say nothing about the category.
  • Single runs. One screenshot is an anecdote. One week is not a pattern.
  • Blended metrics. A combined “visibility score” that adds mentions to links hides which one moved.
  • Causal claims from correlation. A citation gain after a page change is an association. Engines, models, and competitors change in the same window. Claim causation only with a control group of untouched pages, or a repeated effect across several changes.
  • Unrecorded changes. Log every site release, model update you know of, and tracking outage against the dates.
  • Mixing panels. Changing engines, settings, or the query set mid-window breaks the trend. Start a new baseline instead.

Sources

Bring us the domain.

We will tell you where you are absent, which page is in the way, and what has to be true before a citation is possible.

Request an audit