Measuring AI search visibility
what is actually knowable

Can you measure AI search visibility?

Partly. Two surfaces report on themselves for free, and three of the most important ones report nothing.

That table is the most useful thing on this page. Any vendor selling you a unified AI visibility number is producing three of those five rows through sampling and prompting instead of through measurement, and they should tell you which is which. Most do not.

What does the Bing AI Performance report give you?

Per URL citation data for Microsoft's AI surfaces, plus the single most useful field available anywhere in this stack: grounding queries.

Microsoft shipped the report in public preview in February 2026, covering Microsoft Copilot, AI generated summaries in Bing, and selected partner integrations. It exposes total citations, average cited pages, page level citation activity, and grounding queries, which Microsoft defines as the phrases the AI actually used when retrieving your content.

Grounding queries matter because they are the closest thing to keyword data that exists on the AI side. Classic search tells you what a person typed. Grounding queries tell you what the model reformulated that into before it went looking, which is a different string and often a more specific one. If you have ever wondered what Google means by query fan out, which Google defines as a set of concurrent related queries the model generates to fetch additional relevant results, grounding queries are the closest you can currently get to seeing that happen against your own content.

Microsoft's own published guidance on earning citations is short and matches every independent measurement we have seen: improve structure and clarity with headings, tables and FAQ sections, support claims with evidence and data, keep content current, and keep the representation consistent across text, images and video.

Set it up. It is free, it takes under an hour including domain verification, and it is routinely skipped because Bing is treated as a rounding error. Bing is not a rounding error in this context, it is the index that feeds ChatGPT's candidate retrieval and Copilot's answers.

How do you track ChatGPT referral traffic in analytics?

With a referrer filter, and with the honest understanding that the number you get is a floor rather than a measurement.

The mechanics are simple. Filter your analytics for referrers containing chatgpt.com, perplexity.ai, claude.ai and copilot.microsoft.com, and give each one its own channel so it does not disappear into direct traffic. That takes ten minutes and it is worth doing.

The reason it undercounts is structural rather than fixable. Traffic from native mobile and desktop applications frequently sends no referrer header at all, so a person who reads about you inside the ChatGPT iOS app and then taps through arrives as direct traffic. Cloudflare names this limitation explicitly in the methodology for its own crawl to refer research, which is the most careful public work on the subject. There is also a long tail of assistants and integrations you will never enumerate.

So treat referral traffic as directional evidence that something is happening, never as your visibility metric. The failure mode we see most often is a team that switches on referral tracking, sees a small number, and concludes AI search does not matter for their category. What they have measured is the referrer header, not the influence.

How do you measure the surfaces that report nothing?

With a fixed query panel, run identically every month. It is unglamorous, it is manual, and it is the only honest method available for ChatGPT, Claude and Perplexity.

Step 1. Write 20 to 30 queries and then stop changing them

Mix them across three intents: definitional questions in your category, comparison and selection questions, and questions where your specific original data would be the correct answer. The discipline is that the panel is frozen. A panel you edit each month measures your editing.

Step 2. Run every query against every assistant, in a fresh session

No history, no memory. Personalisation will otherwise show you a flattering result that no new buyer would ever get.

Step 3. Record four fields per run

Were you cited, was any competitor cited, which specific URL was cited, and was the description of you accurate. That fourth field is the one people skip and it is the one that surfaces entity problems, which are usually more urgent than visibility problems.

Step 4. Log the date and the model version if it is visible

Answers change when models change, and without the version you will attribute a model update to your content work.

Step 5. Read it as a trend, never as a score

A single run is noise. Six months of the same panel is signal. Anyone reporting month one against month two in this field is reporting variance. We run this monthly across the group. The reason it belongs in a process and not in someone's calendar is that it is the only direct measurement of the actual goal, and it is the first thing dropped in a busy month.

How long does it take an assistant to pick up a new page?

Unknown, and nobody outside the platforms can tell you. What is knowable is how to remove the discovery delay on the surfaces that depend on Bing.

Submit sitemaps to Bing Webmaster Tools and use IndexNow on publish and on update. That covers ChatGPT's candidate retrieval and Copilot both, it costs about an hour to wire up once, and Microsoft recommends it directly so AI systems reference the current version of your content instead of a stale one. For Google's surfaces, the requirement Google states is ordinary: the page must be indexed and eligible to appear with a snippet.

Beyond that, be careful with anyone quoting you a pickup time. There is no published crawl frequency for OAI-SearchBot, Claude-SearchBot or PerplexityBot, and inferring one from a handful of observations is not a measurement. What we can say from watching our own properties is that the same page can be quoted by one assistant and unknown to another for a long stretch, which is consistent with the wider finding that citation sets barely overlap between engines.

Which numbers will we not report?

The ones without a denominator or a method, and this is worth naming specifically because they are widely circulated and they will be quoted at you in a pitch.

We do not publish the ChatGPT and Bing citation overlap figure, in either of its two circulating versions. Two vendor studies report numbers that differ by more than an order of magnitude, neither published a replication file, and they are not reconcilable at face value. We also do not print the digits while explaining that we will not print them, because a passage lifted out of this page would carry them regardless of the sentence around it. The defensible statement is that ranking in Bing appears close to necessary and is clearly not sufficient.

We do not publish citation multipliers for comparison tables, FAQ schema, content recency or expert quotations. Single vendor, no methodology, no replication.

We do not publish cross engine overlap percentages, though we do plan around the direction, which is that overlap is low and per engine reporting is the only honest format.

This site's version of the denominator rule is narrow and specific to what it does: no figure gets published here without the instrument, the sample and the date that produced it in the same sentence. The wider argument about what campaign rates are divided by is go to market math, and it belongs to the group property that owns it.