How to measure AI visibility without the engines' data: a sampling method

As of 9 October 2026, none of OpenAI, Anthropic or Perplexity documents a site-level citation report for site owners. Two engines now give you partial data. The rest give you nothing. If you want a number you can track month to month, you have to build it yourself.

This article explains what the engines do report, then sets out a sampling method that works for the engines that do not.

What do the engines report today?

Two first-party reports exist. Both are useful, and both are narrower than they sound.

Bing Webmaster Tools has an AI Performance report. Microsoft’s announcement says it shows how publisher content appears across Microsoft Copilot and AI-generated summaries in Bing. It reports total citations, average cited pages per day and “grounding queries”. Microsoft describes those as “the key phrases the AI used when retrieving content that was referenced in AI-generated answers”.

Read the caveats in the same post. For grounding queries, Microsoft says “the data shown represents a sample of overall citation activity”. It also says page-level counts reflect “how often pages are cited, not page importance, ranking, or placement”.

Google Search Console has a Generative AI performance report. Google’s help page says it includes impressions for AI Overviews and AI Mode. An impression is defined as “how many times links to your site were shown to a user in a generative AI feature on Google Search”. The page describes impressions only. If the report is missing, Google says this “usually happens when your site hasn’t received enough impressions in generative AI features”.

Google’s AI features documentation adds that AI feature traffic is also folded into the main Performance report under the “Web” search type. So you can see totals, but you cannot separate an AI Overview click from a classic result click there.

What do those reports leave out?

Most of what clients ask about. Put the two reports side by side and the gaps are clear.

• As of 9 October 2026, none of OpenAI, Anthropic or Perplexity documents a site-level citation report in the crawler documentation linked above.
• No data on whether an answer mentioned your brand without linking to it.
• No data on whether what the assistant said about you was accurate.
• No data on which competitors were cited alongside you.

Those are the questions clients actually ask. A sampling method is the only way to answer them today.

The sampling method in one paragraph

Write a fixed set of prompts that your buyers would plausibly ask. Run each prompt several times in each assistant on a fixed schedule. Record whether your brand was mentioned, whether your site was cited with a link (mention rate and citation rate), and whether the description was accurate. Report the counts, the run dates and the method. Do not report anything you did not observe.

The rest of this article is the detail.

1. 1

   Prompt set

   Category, vendor and comparison prompts. Wording frozen; changes logged.

2. 2

   Engines and runs

   Each prompt run several times per engine, fresh sessions, dates recorded.

3. 3

   Score each answer

   Mentioned, cited, accurate, competitors cited.

4. 4

   Report

   Counts with sample size, dates and method. Nothing not observed.

Beside it, not added to it

Search Console Generative AI performance report (Google) and Bing Webmaster Tools AI Performance (Copilot).

Step 1: build the prompt set

Start from your service pages and your sales conversations, not from keyword tools. For each service, write prompts in three shapes:

1. Category prompts. “What is [service] and when do I need it?”
2. Vendor prompts. “Which companies offer [service] for [industry] in [country]?”
3. Comparison prompts. “How do I choose between [approach A] and [approach B]?”

Keep the set small enough to run every month without shortcuts. A set you run reliably beats a large set you abandon. Freeze the wording once agreed. Changing a prompt resets its history.

Step 2: decide the engines and the runs

Pick the assistants your buyers use. Run each prompt more than once per engine, because answers vary between runs. Record every run, not just the best one. Use fresh sessions with no personalisation or memory where the product allows it, and note where it does not.

Step 3: score each answer

Score every run on four yes-or-no questions. Keep the rubric this simple so two people would score the same answer the same way.

Field	Record
Mentioned	Was your brand named in the answer?
Cited	Was a URL on your domain shown as a source?
Accurate	Was what it said about you correct, judged against your own site?
Competitors cited	Which other domains were cited? (list)

Accuracy is the field people skip, and it is the one that finds the costly problems. An assistant that describes a retired service, a wrong price or a wrong location is doing damage you will not see in a citation count.

A worked example of the output

Here is how a monthly summary might read for one engine and one prompt set. The figures are illustrative, chosen to show the format. They are not measurements of any real company.

That block is the deliverable. It states the sample size, the dates and the method, so anyone can repeat it. It gives counts rather than percentages because 4 of 36 is a fact, and the percentage invites more precision than 36 runs can support.

Step 4: combine with first-party data where it exists

For Bing and Google, put the first-party numbers next to your sample. The Bing report tells you which pages are cited and for which grounding queries. The Search Console report tells you where AI impressions are growing. Your sample tells you what the answers actually say. Each covers a blind spot in the others.

Do not add the sources together into one score. They measure different things, over different populations, with different sampling.

What this method cannot tell you

Be explicit about limits in every report.

• Your prompt set is not the population of real queries. It is your best guess at them.
• Answers vary by run, by location, by account history and by the day’s model version. A change in your count may be noise.
• Microsoft describes the grounding-query data in its AI Performance report as a sample. Yours is a smaller one.
• A mention is not a visit, and a citation is not a conversion. Attribution needs your analytics, and even then it is partial.

How often should you run it?

Monthly is a sensible default. It is frequent enough to notice an accuracy problem before it spreads, and infrequent enough to run properly. Re-run immediately after a major site change or a rebrand, since those are the moments descriptions go stale.

What we saw on our own properties

The advice in Step 4 not to add the sources together comes from our own data. On two of the eight sites in our case studies, the two first-party reports moved in opposite directions over the same summer:

Site	Google AI-feature impressions	Bing AI citations
bagassepackaging.co.uk	1,667 in June, 975 in September	40 in May, 188 in September
shopifywebdeveloper.co.uk	278 in July, 113 in September	108 in July, 56 in September

The share of Google impressions recorded in Search Console’s Generative AI performance report also varied widely between sites, from 1.4% on shopifywebdeveloper.co.uk to 17.0% on singlecrystaldiamondring.com, over windows starting between May and June 2026. One combined score would have hidden both facts. Neither report says what the answers said about the company, which is what the sample is for.

What we don’t claim

This method produces a tracked, repeatable observation. It does not produce a market share figure, and nothing in it predicts whether any assistant will cite you next month. We do not promise mentions or citations to anyone. We promise a method, a record and an honest statement of what the record can and cannot show.