How to track whether AI assistants cite your content

You track AI citations with four signals, none of which is complete alone: the AI Performance report in Bing Webmaster Tools, the AI Assistant channel in Google Analytics, AI bot hits in your server logs, and repeated prompt sampling. Read them together weekly and treat every number as a floor, not a total.

No tool shows you every citation. The honest goal is a trend line you trust, built from sources that each admit their blind spot.

What can you actually measure today?

You can measure four things: citations reported by a search engine, visits that arrive from an assistant, fetches by AI bots, and how often your brand appears when you ask the questions your buyers ask. Each answers a different question, and each misses something the others catch.

Method	What it tells you	Cost	Blind spot
Bing AI Performance report	Citation counts, cited pages, grounding queries (Copilot, Bing summaries, select partners)	Free	No click data; Microsoft surfaces only
GA4 AI Assistant channel	Visits that arrived with a recognised assistant referrer	Free	Needs a referrer; excludes Google AI Overviews and AI Mode
Server and CDN logs	Which AI user agents fetched which URLs	Free if you keep logs	A fetch is not a citation
Prompt sampling	How often you appear across many runs of a question	Free by hand, paid at scale	Answers vary run to run

Does Bing show you which pages Copilot cites?

Yes. Microsoft launched AI Performance in Bing Webmaster Tools as a public preview on 10 February 2026. It shows total citations, the average number of unique pages cited per day, and the grounding queries the AI used when it retrieved your content. It covers Copilot, Bing’s AI summaries and select partner integrations.

This is the only first-party citation count we found from a major vendor, so verify your site in Bing Webmaster Tools first. Microsoft’s announcement says citation counts do not indicate ranking, importance or placement within an answer, and it describes no click or engagement data.

Use it for two jobs. Sort cited pages to see which of your posts get pulled into answers, and read the grounding queries to see the phrasing that triggers it.

How do you count AI referral visits in GA4?

Google Analytics has a default channel called AI Assistant. Its help page describes it as visits from sources like ChatGPT, Gemini, Deepseek, Copilot or Grok, assigned when the medium is “ai-assistant” and the referrer matches Google’s list. It explicitly excludes Google’s AI Overviews and AI Mode.

Open the Traffic acquisition report, filter to that channel, and add session source and medium as a secondary dimension. That tells you which assistant sent each visit and which pages they landed on.

Two limits matter. First, a visit only counts if the assistant passes a referrer, so some real arrivals will sit in Direct. Second, Google’s list is Google’s list. If an assistant you care about shows up as ordinary referral traffic, build a custom channel group that matches its domain, and review it when assistants change.

Semrush’s April 2026 channel study, which we cited in our post on how long SEO takes, put AI and LLM referrals at 0.14% of all traffic through 2025. A flat line here is normal. A rising line on your best pages is the signal.

What do server logs tell you about AI bots?

Server logs show which AI user agents requested which URLs, and they work even when no human ever clicks. OpenAI documents three that matter here: OAI-SearchBot for its search features, GPTBot for training data, and ChatGPT-User, which visits a page when a user asks ChatGPT or a custom GPT a question.

Of the three, ChatGPT-User is the closest thing to a citation trace in a log. OpenAI’s documentation says it fires on user questions, not automatic crawling. A request from it for your pricing page means a person asked something that sent the assistant there.

OpenAI publishes IP lists for each bot, so you can check that a request claiming to be one is genuine. Filter your logs by user agent, group by URL, and compare week over week. Our generative engine optimization post covers why the three bots play different roles and which robots.txt rules to set. Here the point is only to count them.

A fetch is not a citation. The assistant may read your page and quote someone else. Treat log hits as evidence of retrieval attempts, and use the other signals to confirm what reached an answer.

Can you test it yourself by asking the assistants?

You can, but one query proves nothing. SparkToro and Gumshoe.ai published research on 28 January 2026: 600 volunteers ran 12 prompts through three AI tools, 2,961 runs in all. By their math, the odds that ChatGPT or Google’s AI gives the same brand list twice are under 1 in 100.

So a screenshot of your brand at number two tells you about one roll of the dice. The study’s author recommends running each prompt at least 60 to 100 times and averaging the results. He argues the useful metric is visibility percentage across dozens to hundreds of prompts, and that any tool selling you a fixed ranking position in AI is not measuring something stable.

That caveat applies to the paid tracking tools too. Ask any vendor how many runs sit behind each number. The study is also not peer reviewed, and its authors disclosed a commercial interest in AI visibility tools, so read it as a strong warning, not a law.

A workable manual version: write twenty questions your buyers actually ask, in their words, not yours. Run each one several times a month in the assistants you care about. Record whether you are named, whether your URL is linked, and who else is. Count the share of runs, not the position.

What about Google AI Overviews?

Google tells you where the data goes, not how much of it is yours. Its AI features documentation, updated 10 December 2025, says traffic from AI Overviews and AI Mode is included in overall search traffic in Search Console, reported in the Performance report under the Web search type. The page we read describes no way to separate it.

That matters because Google’s own AI surfaces are the ones you cannot isolate. The GA4 AI Assistant channel leaves them out by definition, and Search Console folds them into ordinary web search. Your GA4 channel is a partial view, and you should say so in any report you write from it.

For Google, watch Search Console impressions and clicks on the pages you expect to be quoted. Pair that with the prompt sampling above, run against the questions where AI Overviews show up.

A weekly routine that takes thirty minutes

Do this on the same day each week and keep the results in one sheet.

1. Export Bing’s page-level citations and note new pages.
2. Pull the GA4 AI Assistant channel by landing page.
3. Count ChatGPT-User, OAI-SearchBot and other AI agent hits per URL from your logs.
4. Run your twenty buyer questions a few times each and log the share of runs that name you.

Then compare pages across the four lists. A page that appears in three of them is working. A page that gets bot hits and no citations probably answers the question too loosely or buries the number.

This is the shape of job we hand to a tomte, our word for one production agent with one defined job: collect the four exports, write the sheet, flag the pages that moved. The judgement about what to write next stays with whoever owns the content plan.

Where this fits in a content engine

Tracking only pays off if something acts on it. The same site that publishes daily should read these signals and feed them back into the queue.

That is how we run content on the inbound machine, our 363-product B2B catalogue site: a daily automated pipeline, with automated gates deciding every publish and quality inbound leads within its first month, as of July 2026. Measurement is the other half of that loop, and it is part of what the content engine install covers.

The plain version

Start with Bing Webmaster Tools and the GA4 AI Assistant channel today, both free. Add a log filter for AI user agents this week. Build a twenty-question prompt list and sample it monthly with many runs per question. Report trends and shares, never single results, and say which surfaces you cannot see.

If you want this set up for your site, send us your domain and the questions your buyers ask. You will get a written plan within one business day. No meetings, fixed price. Start async.