By Zach Wennstedt, founder · Updated November 6, 2026
You can’t get one clean “AI citations” number, and anyone selling you one is sampling something. What you can do is combine five honest signals: AI referral visits in analytics, AI crawler hits in your server logs, a repeatable prompt panel, and the AI reports Google and Bing now give site owners. Together they’re reproducible, and reproducible is what lets you tell a real change from noise.
What can you actually measure?
Five signals, each with a blind spot. Use all of them, because the blind spots don’t overlap.
| Signal | Where it lives | What it tells you | Blind spot |
|---|---|---|---|
| Referral visits | GA4 | People who clicked a cited link | Answers nobody clicked; referrer or UTM tags can be stripped |
| Bot hits | Server logs | Which AI crawlers fetch which URLs | A fetch is not a citation |
| Prompt panel | Your own sheet or script | Whether you’re named or linked for the questions you care about | Answers vary run to run, so it’s a sample |
| Generative AI report | Search Console | Impressions of your URLs in AI Overviews and AI Mode | Impressions only |
| AI Performance | Bing Webmaster Tools | Citations and grounding queries in Copilot and Bing | Microsoft surfaces only |
How do you find AI referral traffic in GA4?
Start with what Google Analytics already does. Its channel documentation lists an “AI Assistant” channel in the default channel group, for sources such as ChatGPT, Gemini, DeepSeek, Copilot and Grok, and says traffic from Google’s AI Overviews and AI Mode is classified as Organic Search instead (Google Analytics Help). That second part matters: GA4 will never show you AI Overview clicks as their own line.
Open Reports, then Acquisition, then Traffic acquisition. Read the AI Assistant row, then switch the dimension to session source / medium to see which assistants send the visits. Don’t stop at the default channel, though. In an Exploration, filter session source with a regular expression like chatgpt|perplexity|gemini|copilot|claude and compare it to the channel total. Anything that shows up in one and not the other is a classification gap you can fix with a custom channel rule.
Why do some AI visits show up as Direct?
Because the click has to carry a referrer or a tag, and sometimes it doesn’t. OpenAI’s publisher FAQ says ChatGPT search adds utm_source=chatgpt.com to links so you can see the visits. Then on October 14, 2026, Search Engine Roundtable reported that after the GPT-6 rollout, links stopped carrying both the tag and the referrer. In one analyst’s sample of about 500 pages, ChatGPT visits fell by about 31,000 a day while Direct rose by about 26,000 a day. The analyst later said it was a bug and that tagging had resumed.
The lesson isn’t about ChatGPT. It’s that referral data is only as stable as someone else’s link behavior. Annotate the date in GA4 whenever you see a step change, and watch one more line: Direct sessions that land on deep, non-brand URLs (not the home page). If that line climbs while the assistant lines fall, you’re probably looking at stripped referrers. That’s an inference, not proof, so treat it as a flag to check the logs, not a number to report.
What do server logs show that analytics can’t?
Which bots touched which pages, including the ones that never send a human. This is the signal analytics structurally can’t see. Here are the user agents to filter on, per each vendor’s own documentation:
| User agent | Purpose, per the vendor | robots.txt |
|---|---|---|
OAI-SearchBot | Surfaces sites in ChatGPT’s search features | Allowing it helps you appear |
GPTBot | Crawls content that may be used to train OpenAI’s models | Controlled independently of OAI-SearchBot |
ChatGPT-User | User-initiated visits, such as fetching a page when someone asks a question | OpenAI says rules “may not apply” |
PerplexityBot | Surfaces and links sites in Perplexity results; not used for foundation model training | Respected |
Perplexity-User | Fetches a page when a user asks a question | Generally ignored, because a user requested the fetch |
Sources: OpenAI’s bot documentation and Perplexity’s. A quick count of what each bot requested, assuming a standard combined-format access log:
grep -E "OAI-SearchBot|GPTBot|ChatGPT-User|PerplexityBot|Perplexity-User" access.log \
| awk '{print $7}' | sort | uniq -c | sort -rn | head -20
User-agent strings are trivially forged, so verify the heavy hitters against the vendors’ published IP lists (OpenAI publishes searchbot.json, gptbot.json and chatgpt-user.json; Perplexity publishes perplexitybot.json and perplexity-user.json). Our log file analysis guide covers the verification step, and what AI crawlers actually read covers what they do with the page once they have it.
How to read it: OAI-SearchBot and PerplexityBot hits tell you the page is available to those search products. A ChatGPT-User or Perplexity-User hit is the interesting one, because a person’s question triggered a live fetch of that URL. That’s the closest thing in your logs to “this page was considered for an answer.” It still isn’t proof the page was cited.
How do you build a prompt panel?
A prompt panel is a fixed list of questions you ask the same assistants on a schedule, with every answer logged. It’s the only method that measures the thing you actually care about: does your site get named or linked when a buyer asks.
- Pick 20 to 30 prompts from real customer questions: calls, form fills, sales emails, Search Console queries. Split them into brand (“is [your company] any good?”), category (“best [service] in [city]”) and comparison (“[you] vs [competitor]”).
- Run each prompt three times per surface, per week. Output varies between runs, so a single answer is an anecdote.
- Hold conditions steady: same surface, signed-out or the same account state, same location, and a note on which model or mode you used.
- Log one row per answer in a sheet or CSV.
date,surface,prompt_id,run,cited_domain,linked,position,competitors_cited,answer_url
2026-11-06,chatgpt-search,cat-07,1,yes,yes,2,"acme.com;bolt.io",https://...
Your headline metric is citation rate: answers that cite your domain divided by answers collected. How much should you trust it? With 25 prompts times 3 runs, that’s 75 answers per surface per week. At a true rate of 20%, basic binomial math puts the 95% margin at about plus or minus 9 points (1.96 × √(0.2 × 0.8 ÷ 75)). The runs aren’t perfectly independent, so the real margin is wider. A jump from 20% to 24% in one week is noise. A move from 20% to 45% held over four weeks isn’t.
Record whether you used the consumer app or an API. They can answer differently, and mixing them silently will wreck your trend line.
What do Search Console and Bing report now?
Google added a dedicated view. The generative AI performance report was announced on June 3, 2026 and, per Google, reached all sites worldwide as of August 31, 2026. It reports impressions only, meaning how often URLs from your site appeared in AI Overviews and AI Mode, with tabs for pages, countries, devices and dates, and a split between text-based and multimodal searches. The help page adds three limits worth knowing: the usual 1,000-row cap applies, Search Labs experiments aren’t included, and if you don’t see the report it may mean your site hasn’t had enough impressions in these features (or you’ve opted out of them).
For clicks, Google’s AI features documentation says AI Overviews and AI Mode are counted in the overall Search Console data, in the Web search type. So you can see how many times you appeared in AI answers, and you can see total clicks, but the documentation describes no separate AI click report. The same page says there are no additional technical requirements to appear: the page must be indexed and eligible to show with a snippet. If you want the Google-specific tactics, see our AI Overviews SEO page.
Microsoft’s version is AI Performance in Bing Webmaster Tools, announced February 10, 2026. It shows total citations, cited pages and “grounding queries” (the phrases the AI used when retrieving cited content, which Microsoft says are a sample) for Copilot, Bing’s AI summaries and select partners. It doesn’t cover ChatGPT, Perplexity or Google.
How do you put the signals together?
One weekly readout, five lines, same order every time.
| Metric | Source | Act when |
|---|---|---|
| AI Assistant sessions, plus regex-matched sources | GA4 | The four-week trend moves; check Direct on deep URLs first |
| Bot requests by URL, verified IPs only | Server logs | A priority page gets zero search-bot hits |
| Citation rate by prompt bucket | Prompt panel | A bucket moves more than the margin for four weeks |
| AI Overview and AI Mode impressions by page | Search Console | Priority pages are absent from the report |
| Citations and grounding queries | Bing Webmaster Tools | Grounding queries reveal a topic you don’t cover |
When you change a page, write the date next to it and wait four weeks before judging. Change one thing at a time where you can, or you won’t know what moved the number.
What should you not measure?
- A single screenshot. One good answer proves it can happen, not how often.
- A tool’s “AI visibility score” you can’t audit. Ask how many prompts, how many runs, which surfaces and which location. If the vendor won’t say, you can’t tell signal from noise.
- llms.txt as a scoreboard. A file existing tells you nothing about whether anyone cites you.
- Raw bot hits as “citations.” A crawl is access, not endorsement.
If you’d rather have this instrumented than build it yourself, our technical SEO work includes the GA4 rules, log parsing and a prompt panel you own.
