How to Measure Your Website's AI Visibility: What Google, Bing, Yandex and GA4 Actually Show — and What No Tool Will

The honest answer is an inconvenient one: there is no single AI visibility number. There are several official reports — from Google, Bing, Yandex and in GA4. Each one sees its own piece of the picture, and each has blind spots. And some of what you want to know, nobody can show you today.

Below: what each tool actually shows, how to set it up, how to calculate the metrics without fooling yourself, and where the limits of what we know lie. With data for Moldova, a real case from my practice, and the statistics that usually get left out. All facts were checked against the platforms' official documentation as of October 10, 2026.

Short answer#

A site's AI visibility isn't measured with one metric. It takes a set of six data sources:

• the Generative AI performance report in Google Search Console — impressions in AI Overviews and AI Mode, with no clicks or queries;
• the AI Performance report in Bing Webmaster Tools — citations in Copilot;
• "Site visibility in Alice AI" in Yandex Webmaster — share of voice (SoV);
• the AI Assistant channel in GA4 — visits from ChatGPT, Gemini, Claude and other chatbots;
• server logs — AI bot visits;
• your own prompt tracking — the share of answers from ChatGPT, Perplexity and other AI systems that mention the brand (only with a confidence interval).

No tool shows the queries for which Google surfaces your site in AI answers, separate clicks from AI Overviews and AI Mode, or impressions in ChatGPT and Perplexity. Any share calculated from manual prompt checks needs a confidence interval — otherwise it's misleading.

Why there is no single AI visibility number#

In AI search, a site's path to the user has several stages: a bot crawls the page → the page gets indexed → the AI system selects it as a source → the user sees the link → the user clicks through → the user converts. Each tool sees only part of this chain.

Rule: you can't add up or directly compare numbers from different rows of this table. A 30% Citation Share in Bing and a 30% SoV in Yandex are shares of different things. For every metric, you need to know the denominator.

Context: what the research says — and where it falls short#

Before you crunch your own data, it helps to understand the scale of the shift. Below are the most-cited studies, with the caveats that usually go unmentioned.

An important caveat: Yandex also cites the Clarity study in its help page on Alice AI. But the sample there is media sites only, and the metrics are sign-ups and subscriptions that Clarity detects automatically (smart events). You can't carry "AI traffic converts 3x better" over to an online store or a service business without testing it on your own data.

Moldova by the numbers: which search engines and AI systems matter#

In a local market, platform shares determine your toolset. According to StatCounter, in September 2026 Google held 90.64% of search in Moldova, Yandex 6.57% and Bing 1.7%. For comparison, Yandex has 0.32% in Romania.

The takeaway for a Moldovan site: Google Search Console is a must; Yandex Webmaster if you have a sizable Russian-speaking audience; Bing Webmaster Tools for the Copilot data, and because setup takes a couple of minutes via import from GSC.

For AI chatbots, the picture looks like this: in September 2026, ChatGPT accounted for 81.86% of referrals from AI interfaces to websites, Google Gemini for 12.25%, Perplexity 2.95%, Claude 1.86% and Microsoft Copilot 1.09% (StatCounter). Over 12 months (September 2025 – September 2026): ChatGPT 83.11%, Gemini 6.8%, Perplexity 5.63%, Copilot 2.71%, Claude 1.75%

In September, Gemini was at almost double its 12-month average.

How to read this correctly. StatCounter doesn't measure how many people use chatbots. It measures the share of referrals from chatbots to sites that run its tracker. That means two things:

• "ChatGPT — 82%" doesn't mean 82% of people in Moldova use ChatGPT. It's the share among people who clicked through to a site from a chatbot link.
• A platform that rarely gives links will look smaller than its actual audience.

Most importantly, a national benchmark is not your site. In the case below, ChatGPT accounts for 97% of AI sessions, not 82%.

Layer 1. Google Search Console: the Generative AI performance report#

What this report is#

Since August 31, 2026, the report has been available to all sites worldwide (Google help). It shows impressions of links to your site in AI Overviews and AI Mode. Discover has a separate, similar report.

• Impression — the number of times a link to your site was shown to a user inside an AI feature.
• Dimensions: pages, countries, devices, dates. There's a search type filter: text-based or multimodal (Google Lens, Circle to Search, image search).
• The report has no clicks and no queries.
• The same data is already included in the regular "Performance" report (search type "Web"). The separate report simply breaks out the AI portion.
• Search Labs experiments are not included.
• Dates are in Pacific Time (PT). The usual Search Console limits apply, including 1,000 rows per table.

Don't see the report? According to Google's help, this usually means the site doesn't have enough impressions in AI features. The second reason is that the site has been excluded from generative AI features via a setting in Search Console. For small Moldovan sites, a missing report is normal, not an error.

How Google counts position and clicks in AI answers#

This matters for reading the main report, which does include clicks and position (Google help):

• An AI Overview occupies a single position in the results, and every link inside it gets that same position.
• AI Mode calculates position the same way as regular search results. A follow-up question in AI Mode counts as a new query.
• In both AI Overviews and AI Mode, a click on a link to an external page counts as a click.

The practical implication: clicks from Google's AI answers are already in your main report, but you can't separate them from regular clicks.

Step by step#

1. Search Console → "Performance" → expand it and open the Generative AI performance report.
2. Select a 3-month date range and group by week — daily data is noisy.
3. Export the "Pages" tab.
4. Export the main "Performance" report by page for the same period.
5. Join the tables on URL and calculate the AI impression share.

Don't be surprised if the chart total doesn't match the sum of the table: the chart counts impressions by property (several links to the site in one answer count as one impression), while the pages table counts each URL separately (Google help).

Formula:

Page AI impression share = impressions in the Generative AI performance report ÷ all page impressions in the Performance report
Wilson interval = (p + z²/2n ± z·√(p(1 − p)/n + z²/4n²)) ÷ (1 + z²/n)

The comparison is valid because AI impressions are a subset of total impressions. Pull both values grouped by page. The metric shows which pages "live" in AI answers. Pages with a high AI impression share and a falling CTR in the main report are the first candidates for analysis.

Case: 684,000 impressions in 3 months#

Over 92 days (July 7 – October 6, 2026), the site got 684K impressions in AI Overviews and AI Mode, an average of about 7.4K per day. From late August to early October, the daily average roughly doubled: on the chart, about 5K impressions per day in the last week of August and about 10–11K in the first week of October (peaking at 11,215 impressions on October 6). These values are read off the chart by eye, so treat them as an order-of-magnitude estimate, not an exact percentage.

What we can claim: the site's presence in Google's AI answers is growing. What we can't claim: that our work caused the growth. In late September, industry trackers recorded a sharp rise in the share of branded queries showing AI Overviews — according to DemandSphere, from ~26% to over 80% of tracked queries — and Google didn't announce the change (Search Engine Land). The growth on the chart may come from Google itself expanding AI features rather than from improvements to the site. Without a control group, you can't draw a causal conclusion.

Layer 2. GA4: the AI Assistant channel#

What changed in GA4#

GA4's default channel group now includes an AI Assistant channel. According to the official Google Analytics help:

• the channel includes visits from ChatGPT, Gemini, Deepseek, Copilot, Grok and other AI assistants;
• if the referrer matches the list of AI assistants, GA4 assigns the medium ai-assistant and the campaign (ai-assistant);
• visits from AI Overviews and AI Mode are not included in this channel — they're counted under Organic Search.

The last point is the key one. You can't isolate Google's AI traffic in GA4: it's blended into organic. The AI Assistant channel shows only visits from chatbots.

Case: what the AI Assistant channel showed#

The "whole site" values are reconstructed from the percentages GA4 shows in the totals row (percent of total and difference from average), so they're approximate.

What the data shows:

1. ChatGPT drives 97.7% of AI sessions (6,437 of 6,589) and the same share of the channel's key events (775 of 793). For comparison: across Moldova in September, ChatGPT accounted for 81.86% of AI referrals and Gemini for 12.25% (StatCounter), yet on this site Gemini has just 1.5%. Country-wide averages and an individual site's traffic mix can diverge sharply, so decisions should rest on your own data.
2. AI traffic is noticeably more engaged than average. The engagement rate is 54.6% vs. ≈32.4% site-wide (95% interval for the AI channel: 53.4–55.8%), and engagement time is 54 sec vs. ≈37. Events per session, though, are slightly lower (10.3 vs. ≈10.9), so "better across the board" is an overstatement.
3. 0.31% of sessions produce 0.51% of key events. That's 12.0 per 100 sessions vs. ≈7.3 site-wide — about 1.6x more. Allowing for GA4's rounding of percentages, the ratio is between 1.60 and 1.69.
4. Small sources are noise, not findings. claude.ai shows 73% engaged sessions and almost 6 minutes per session, but that's 11 of 15 sessions: the interval runs from 48% to 89%. copilot.com has 4 sessions and not one engaged session. You can't rank assistants on rows like these.
5. Reconcile source name variants. Here, all Gemini traffic came in as gemini.google.com, but in other exports from my practice I've also seen the source gemini without a domain. Merge the variants before calculating shares — otherwise you understate that assistant's share.

How to interpret this honestly. The 95% confidence interval for 793 events across 6,589 sessions is 11.3 to 12.8 per 100 sessions. Even the lower bound is well above the site average (≈7.3), so the difference is unlikely to be random. But there are three caveats:

• A key event is not the same thing as a converting session: one session can produce several events. The interval here is an approximation.
• It's too early to draw conclusions about individual assistants. Claude delivered 15 sessions and 0 key events — for 0 events across 15 sessions, the 95% interval is 0 to 25 per 100 sessions. Perplexity: 9 events across 38 sessions, with an interval of 11 to 45. At these volumes, you can't compare them with ChatGPT (about 11–13 per 100 sessions). These are Poisson intervals: key events are a count, not a share of sessions, and at small numbers the Wilson formula understates the spread.
• This is correlation, not causation. People who came from ChatGPT may have already known the brand or been looking for a specific offer.

One site, two reports: why you can't divide one by the other#

Screenshots 1 and 2 show the same site over the same period: July 7 – October 6, 2026. Search Console shows 684K impressions in AI Overviews and AI Mode; GA4 shows 6,589 visits from AI chatbots. It's tempting to divide one by the other and get an "AI visibility CTR" of about 1%. But that number means nothing:

• These are different systems. The impressions are Google: AI Overviews and AI Mode. The visits are ChatGPT, Gemini, Perplexity, Claude and Copilot. 97.7% of the visits came from ChatGPT, which doesn't appear in the Search Console report at all.
• Clicks from Google's AI answers don't land in the AI Assistant channel. GA4 counts them under Organic Search, where you can't separate them from regular clicks.

The takeaway: the site has two different AI funnels, and each one has to be measured separately. Google impressions have roughly doubled since late August, but no report will tell you how many of them turned into visits. The closest available signal is the CTR of the same pages in the main "Performance" report: if AI impressions are rising and CTR is falling, that's a reason to dig in. But it's a signal, not proof.

Step by step: setup and validation#

1. Check the default channel. "Reports" → "Acquisition" → "Traffic acquisition" → the AI Assistant row. For a breakdown, switch the dimension to "Session source / medium" and filter for ai-assistant.
2. Check "Unassigned". ChatGPT automatically adds the utm_source=chatgpt.com parameter to links (OpenAI help). If the referrer isn't passed, the visit can end up in "Unassigned" or "Direct". Filter for the source chatgpt across all channels.
3. Create a fallback channel group. "Admin" → "Data display" → "Channel groups" → create a new group. Add an "AI" channel with the condition: Source matches regex .*(chatgpt|openai|gemini|claude|perplexity|copilot|deepseek|grok|meta\.ai|mistral).*. Place it above the "Referral" channel; otherwise Referral will claim the traffic first. The .*….* wrapper is needed because GA4 regular expressions match the entire value by default: without it, chatgpt.com or gemini.google.com may not land in the channel.
4. Test it live. Open "Reports" → "Realtime", visit your site from a link in ChatGPT, Gemini and Perplexity, and see which source each visit shows up under.

Layer 3. Bing Webmaster Tools: citations in Copilot#

Since February 2026, Bing Webmaster Tools has offered an AI Performance report, currently in Public Preview (Bing blog). It covers Microsoft Copilot, AI summaries in Bing and some partner integrations.

Microsoft explicitly calls Citation Share an observational metric: it doesn't show competitor domains, doesn't reflect share of traffic and isn't a ranking (Bing blog, June 2026). The report shows no clicks or visits — only citations.

Step by step: sign in to Bing Webmaster Tools, add your site by importing it from Google Search Console, and open AI Performance a few days later. Export the pages and group the grounding queries by topic.

A useful metric — concentration: the share of citations that comes from the top 10 pages. If 90% of citations land on 2–3 pages, the site's visibility rests on them, and losing one of them will noticeably hit the overall number.

Layer 4. Yandex Webmaster: visibility in Alice AI#

The "Site visibility in Alice AI" tool («Видимость сайта в Алисе AI») launched on April 7, 2026. Here's what the official help says about it:

• data covers the last 3 months and is updated once a week;
• statistics are collected only for queries where the site already ranks high enough;
• the main metric is Share of Voice (SoV): the share of queries in which Alice AI mentioned your site in its answer, out of all queries Alice AI answered;
• the report has no clicks or impressions — according to Yandex, the tool assesses representation, not quantitative metrics;
• for frequently mentioned sites, it shows a range relative to other sites: top 3, top 10, top 20;
• it lists sample queries and sites in the same niche that often appear in answers.

How to read SoV correctly. The denominator isn't every query in the market — only the queries where you already rank well. So a 40% SoV for a small site and a 40% SoV for a niche leader are different things: the leader's denominator is much broader.

One more detail from the help: Alice AI doesn't rank its source links — they're ordered by the logic of the answer (Yandex help). "First place in the AI answer" means nothing here.

If you see "Not enough data yet" («Пока недостаточно данных»), Yandex lists the possible reasons: few queries for which the site appears in search results; low rankings; little text content; Alice AI rarely answers in your niche.

Layer 5. Server logs: which AI bots read your site#

Logs are the only source that sees every AI system, including those that give site owners no reports. But logs show only bot visits, not whether a person saw the answer.

Step by step: export 30–90 days of access logs and keep only the bot lines. As a side benefit, this strips out visitors' IP addresses:

zgrep -hEi "GPTBot|OAI-SearchBot|ChatGPT-User|ClaudeBot|Claude-SearchBot|Claude-User|PerplexityBot|Perplexity-User|Googlebot|bingbot|YandexBot|YandexAdditional" /var/log/nginx/access.log* > ai_bots.log

For bots that publish their IP ranges, check the IPs in the log against the official JSON file. User agents are easy to spoof.

Log-based metrics:

• Coverage = unique URLs visited by an AI search bot ÷ indexable URLs on the site.
• Freshness lag = days from a page update to the search bot's first visit.
• ChatGPT-User and Perplexity-User visits — an indirect sign of which pages AI opens in response to people's requests. This is a proxy, not a visibility metric.

Layer 6. Your own prompt tracking#

ChatGPT, Gemini and Perplexity offer no official visibility reports. That leaves asking them yourself. And here the most important thing is understanding what to count.

The SparkToro and Gumshoe study found that when the same prompt is repeated, the list of brands in the answer matches less than 1 time in 100, and matches in the same order less than 1 time in 1,000. "A brand's position in an AI answer" is noise. The share of answers in which the brand appears at all, however, is noticeably more reproducible over a large number of runs.

Minimum protocol:

1. Lock down the conditions: platform, mode (with or without web search), language, country, date. Start every run in a new chat.
2. Write 10–20 questions your customers actually ask, in Russian and Romanian.
3. Run each question many times. For how many, see the next section.
4. For each answer, record whether the brand is mentioned, whether there's a link to your site, and which domains are cited. Store mentions and links separately.
5. Calculate the share of answers that mention the brand — always with an interval.

Statistics without fooling yourself#

This is the section people skip most often — and that's why most AI visibility numbers on the market mean nothing.

How many runs you need#

If a brand is mentioned in 30% of answers, with a small number of runs the true share could be almost anything (95% Wilson interval):

Why Wilson and not the textbook "±"#

The classic formula "share ± 1.96 × standard error" (the Wald interval) breaks down on small samples and on shares close to 0 or 100%. Brown, Cai and DasGupta showed that its actual coverage behaves erratically and that textbook assurances about when it's safe to use can't be trusted; for small samples, the authors recommend the Wilson or Jeffreys interval.

The Wilson interval formula, where p is the observed share, n is the number of runs, and z = 1.96 for 95%:

Wilson interval = (p + z²/2n ± z·√(p(1 − p)/n + z²/4n²)) ÷ (1 + z²/n) 

Two examples where the difference jumps out:

• The brand isn't mentioned in any of 10 runs. Wald gives an interval of 0–0%, i.e., "definitely never." Wilson gives an honest 0–28%.
• The brand is mentioned once in 20 runs. Wald gives a lower bound of −4.6% — a negative share, which is absurd. Wilson gives 0.9–23.6%.

One more detail: the Wilson interval is asymmetric, so writing "42% ± 8%" is incorrect for it. Report the actual bounds instead: for example, 63 mentions in 150 runs is "42% (95% CI: 34–50%)".

When a "visibility drop" is just noise#

Say the brand was mentioned in 45% of answers last month and in 40% this month. What does that tell you?

To detect a 5-percentage-point change (from 40% to 45%) with 80% probability, you need about 1,530 runs per period. For a 10-point change, about 390. Most manual checks and many monitoring tools work with samples that simply can't detect changes like these.

One more caveat nobody talks about#

All the calculations above assume independent observations. Repeats of the same prompt on the same day, from the same account, from the same country are correlated. So the real uncertainty is greater than the tables show. Spread your runs across days and phrasings.

What no tool will show you#

As of October 2026, no official tool shows:

• the queries for which Google showed your site in AI Overviews or AI Mode;
• clicks from Google's AI features separately from regular clicks — they're blended into the main Search Console report and into Organic Search in GA4;
• impressions in ChatGPT, Gemini and Perplexity — these platforms don't provide reports to site owners;
• unlinked brand mentions — if an AI named your company but cited someone else's site, your reports won't tell you;
• "dark" AI traffic — a person read an answer, then typed your brand name into search or entered your URL directly. In analytics, that shows up as branded search or a direct visit.

You can estimate AI's impact on branded demand only indirectly — through trends in branded queries in Search Console and direct visits in GA4. You can't prove it.

A 30-minute monthly check#

• Search Console: Generative AI performance report impressions by week; top 20 pages; AI impression share for key pages.
• Search Console: CTR of key pages in the main report — is it falling while AI impressions grow?
• Bing Webmaster Tools: citations, concentration in the top 10 pages, new grounding topics.
• Yandex Webmaster: SoV trend and sample queries where competitors are mentioned and you aren't.
• GA4: AI Assistant channel sessions and key events; check "Unassigned" for chatgpt.
• Logs: OAI-SearchBot, PerplexityBot and Claude-SearchBot visits; pages the bots don't visit.
• An interval for every share. For every change, ask: "Is this bigger than noise?"
• Annotations in Search Console and GA4 for every site change — otherwise, three months from now you won't remember what coincided with the growth.

Common interpretation mistakes#

• "We were cited, so people came to us." A citation, a click and a recommendation are three different events. According to Pew, people clicked links inside an AI summary in 1% of visits.
• "We're number one in ChatGPT's answer." The order of brands in answers changes on almost every run. One screenshot proves nothing.
• "The AI Assistant channel in GA4 is all of our AI traffic." No: visits from AI Overviews and AI Mode sit in Organic Search, and some chatbot visits land in "Direct" and "Unassigned". The channel shows a lower bound.
• Universal thresholds like "60% mention rate means strong presence." The share depends on how broad the question is and how many brands compete in the niche. You can only compare against yourself over time and against competitors on the same set of questions.
• "Visibility went up after we added llms.txt." Google's official guide states plainly that its AI features require no llms.txt, no special markup and no chunking of content. For other platforms, no effect has been proven. Coinciding in time is not proof.
• Comparing percentages with different denominators. Yandex SoV, Bing Citation Share and the mention rate from your own tracking are three different quantities.

Bottom line#

AI visibility can't be measured with a single number. It's assembled from six sources (at least for now), each with its own denominator and its own blind spots.

Google shows where you appeared, but not for which queries or how many people clicked through. GA4 shows visits from chatbots but hides Google's AI traffic inside organic. Manual checks without statistics mostly show noise.

Honest measurement isn't about the biggest number. It's about a number with an interval.

Methodology and disclosure#

• All platform facts were checked against official documentation from Google, Microsoft, Yandex, OpenAI, Perplexity and Anthropic on October 9–10, 2026. Interfaces and reports change, so check the sources.
• Confidence intervals were calculated with the Wilson method (95%), proportions were compared with a two-sample z-test, and sample sizes were estimated for 80% power at a 5% significance level.
• Case data comes from projects in my practice and is anonymized. The "whole site" figures in the GA4 case are reconstructed from percentages in the interface and are approximate.
• This article contains no affiliate links and no recommendations of paid tools.
• Disclosure: I provide AI search optimization services. That's why I deliberately relied on official sources and open methodologies, and said so plainly wherever the data is weak.