AI Search Visibility: 542,615 Impressions, 7 Clicks

Between 19 March and 11 August 2026, two pages on this site drew 542,615 impressions from Google at an average position of 3.60. They produced seven clicks. Not seven thousand, not seven hundred. Seven.

That gap is the entire problem with AI search visibility measurement in one line. Every metric a dashboard can show you was green: high impressions, top-three position, growing month over month. The metric that pays salaries was flat. If you had been reporting visibility from that dashboard, you’d have reported a breakout quarter.

What follows is the metric set worth tracking, the reason first-party logs and simulated prompt tracking answer different questions, and what our own Search Console data looks like when you read it honestly. The numbers below are ours. You can check the method; you can’t check the data, which is exactly the problem this article is about.

What AI Search Visibility Measurement Actually Is

AI search visibility measurement is the practice of tracking how often AI systems cite, mention, or recommend your brand in generated answers, and connecting that presence to demand you can bank. It replaces a single ranking number with a set of partial signals, none of which is complete on its own.

The shift matters because the unit of competition changed. In classic search you competed for a position on a page of ten blue links, and the position was observable by anyone. In AI search you compete to be one of a handful of sources a model draws on when it composes an answer, and that composition is private, personalised, and different on every run.

So “are we visible?” stops being one question. It becomes four: are we cited, how often relative to competitors, does any of it reach us, and what happens to the citations that don’t.

The Four Metrics Worth Tracking

Four metrics cover the useful ground. Every vendor framework you’ll encounter is a renaming or a bundling of these, usually with the vendor’s own product mapped onto each row.

The citation gap is the one most measurement programmes skip, and it’s the one that would have caught our seven-click quarter on day three. It’s a subtraction, not a product feature, which is why no dashboard leads with it.

Defining the citation gap

Treat the result as a direction, not a decimal. The second term is always undercounted because branded and direct sessions absorb delayed influence, so the true gap is narrower than the raw arithmetic suggests. It’s still the most useful number in the set, because it’s the only one that goes up when you’re being read and ignored.

Two cautions before you put any of these in a board deck. Published citation benchmarks vary wildly by sample, engine and query intent, and several of the most-quoted figures in this category trace back to no method at all, so check the denominator behind any citation statistic before you compare yourself to it. And where third-party sources dominate the citation set, your own domain is not the lever; the sites the models draw on are, which is a different piece of work from auditing which review sites the assistants actually cite in your category.

First-Party Logs vs Simulated Prompt Tracking

These two evidence sources answer different questions, and the single most common measurement error is treating them as interchangeable readings of one underlying truth.

First-party logs are a census. Search Console, your server logs, and your analytics record events that actually happened: a link was shown, a bot fetched a page, a session arrived. Nobody sampled anything. The limitation is coverage, since a citation that produced no impression and no click leaves no trace in your logs at all.

Simulated prompt tracking is a sample. A tool runs a fixed list of prompts against each model on a schedule and records whether you appeared. This is the only way to observe citations that never generate a visit. The limitation is that you’re sampling a system which re-rolls its answer every time, personalises by account and history, and changes underneath you without notice.

When to use which

Use first-party logs when the question is about outcomes: did this reach anyone, did it convert, is the gap widening. They’re free, they’re yours, and they’re admissible.

Use simulated prompt tracking when the question is about presence: are we in the answer set at all, and who’s there instead of us. Accept it as directional, and hold the prompt list frozen so period-to-period movement means something.

Avoid either alone when you’re deciding budget. Logs alone will tell you AI is irrelevant right up until it isn’t. Simulations alone will tell you you’re winning a race nobody entered.

If you’re buying the simulation layer rather than building it, the platforms in this category differ mainly on prompt-set control, engine coverage and how they handle re-rolls, which is covered vendor by vendor in our review of the GEO platforms that track citations at prompt level.

What 542,615 Impressions and Seven Clicks Look Like

Here’s the worked example. These are the two pages, straight from our Search Console export covering 19 March to 11 August 2026.

Position 3.60 across better than half a million impressions is, on paper, the best-performing content on the site. The click-through rate is 0.001%. For comparison, our homepage converts 1,704 impressions into 43 clicks at position 4.6, which is a normal human CTR of 2.5% at a worse position.

The obvious objection is that we’re cherry-picking a flattering anomaly. The data says the opposite. Strip those pages out and the site’s numbers move like this.

Removing the AI traffic costs 16 clicks out of 954 and makes our average position ten places worse. The synthetic queries were sitting at position three and flattering the average the whole time. Any report built on “impressions” and “average position” was describing a property that didn’t exist.

This is what the citation gap looks like when it’s real rather than theoretical. It is not a rounding error or a tracking bug. It’s the dominant fact about the property, and no visibility dashboard on the market would have surfaced it, because every one of them treats impressions as the numerator of something good.

The other half: what actually arrived

Impressions are one side of the gap. Here’s the referral side, from our analytics over 30 June to 27 July 2026, a 28-day window carrying 1,182 sessions in total.

Sixty sessions from AI assistants against 1,182 total, or 5.08%. Set beside 542,615 impressions, that is the gap stated plainly. But look at the right-hand columns before writing the channel off: ChatGPT sessions engage at 57.14% against Google organic’s 56.76%, and Gemini sessions run nearly twice as long as an organic visit.

So the honest read is not “AI traffic doesn’t convert.” It’s that AI traffic is rare and good, while AI impressions are abundant and worthless. Those are two different findings, and a single blended visibility score would have hidden both. Note also that Perplexity does not appear at all in this window, and that 11.2% of our sessions carry no usable source, so every figure in that table sits on a base with a known hole in it.

Query Length Is the Clearest AI Signal You Already Own

Read the query-length distribution in your Search Console export before you buy any tool. It’s the cheapest AI-detection method available and it uses data you already have.

Humans type short. Machines decompose one question into many long, fully-formed sub-questions and fire them in parallel. That difference shows up cleanly in the export. Ours splits like this across the 1,000 rows Google returns.

A quarter of the queries produced nearly twice the impressions and about 6% of the clicks. That single table tells you more about where your impressions are coming from than most paid dashboards will.

The pattern gets more obvious at the extremes. In our export, 148 of the 1,000 queries are near-identical paraphrases of one narrow idea, together drawing 1,353 impressions and zero clicks. They read like this:

“text matching revenue brackets 0-1 million to 200+ million” · “mapping revenue bands to numeric scores guidelines and best practices” · “first matching condition scoring algorithm revenue ranges” · “points system by revenue range first matching condition”

Nobody types four versions of the same question. That’s decomposition. And in our 12 August internal audit we recorded 26 queries on a single day running over 100 words, the longest at 207 words, opening “i am a 25-34 or 35-44 year old in the consumer goods, food & beverage, retail, or technology industry.” That’s a pasted persona prompt, not a search.

What the Platform Dashboards Will and Won’t Show You

Two search platforms now report AI citation data directly to site owners, and they hand you very different things. Knowing which gives what decides how much tooling you actually need to buy.

Google: impressions without clicks

Google shipped a dedicated Generative AI performance report in Search Console in June 2026, and its limits define what in-house measurement can and cannot do right now.

What it gives you: impressions for AI Overviews and AI Mode, broken out by page, country, device and date. What it does not give you, per Google’s own documentation for the report, is click data. Search Labs experiments are excluded, and the report is still rolling out, so properties below an impression threshold won’t see it yet.

That single omission is the constraint the whole discipline runs into. No clicks means you cannot calculate an AI click-through rate from Google’s own tooling: you get an AI-surface impression count in one report and a blended click count in another, with no key to join them. The citation gap stays measurable in direction, never in precision.

There’s a second subtlety worth knowing. Impressions in this report are aggregated by property, so if two of your URLs appear in the same generated answer, that counts as one impression, not two. A page-level view counts them individually. The two numbers won’t reconcile, and neither is wrong.

Meanwhile the main Performance report still folds AI Overviews and AI Mode into the “Web” search type with everything else, and you cannot filter to isolate them. That blending is why our 542,615 impressions looked like ordinary search performance for months.

Bing: citations, grounding queries, and page-level mapping

Bing gives you what Google withholds. The AI Performance report in Bing Webmaster Tools, in public preview since 10 February 2026, covers Microsoft Copilot, AI-generated summaries in Bing, and selected partner integrations.

It reports total citations, average cited pages per day, and citation counts for individual URLs. Most usefully, it reports grounding queries: the phrases the model actually used when retrieving your content. A June 2026 expansion added citation share, competitor comparison, and query-intent classification.

Read that list against the Google one. Bing tells you which query grounded which page. Google will not join those two dimensions at all, which is the same wall we hit trying to attribute our own long queries to either page. If you want to know what an AI system was looking for when it found you, Bing is currently the only first-party place to see it.

The caveat is proportion. Bing’s slice of total search is small, so treat its citation data as a high-resolution sample rather than a census of AI behaviour everywhere. It’s a leading indicator with real detail, not a market-wide number, and it should never be reported as one.

How to Build Your Own Citation-Gap Baseline

To build a defensible baseline, run five steps in order and write the numbers down before you change anything. The point of a baseline is that it’s boring and repeatable, not that it’s clever.

When to buy the prompt layer instead

Step five is where most in-house programmes stall, and it’s the honest case for buying rather than building: running a frozen prompt set across four engines on a schedule is tedious, and the tedium is what makes the numbers comparable. If you’re outsourcing it, the question that separates real practitioners from resellers is whether they’ll show you a baseline and a method before a proposal, which is the first thing to press on when evaluating an answer-engine specialist’s measurement model. The same test applies to integrated teams selling AI visibility alongside classic SEO, where the two practices are often sold as one retainer but measured with only one of them instrumented.

What You Still Cannot Measure

Honest measurement means naming the boundary. Four limits are structural right now, and no vendor has solved them regardless of what the sales deck implies.

You cannot identify which AI system generated an impression. Search Console does not disclose it. Our own fan-out event is consistent with AI query decomposition, but we cannot exclude a scraper, a rank-tracking tool, or a Google test. We report it as inferred, because that’s what it is.

You cannot join queries to pages. The Queries and Pages exports are separate aggregates with no mapping between them. You cannot prove which page a given long query landed on without live API access, so any tool claiming that link from a standard export is estimating.

You cannot see citations that produced no click. This is the whole reason simulated prompt tracking exists, and it’s why the sampling caveat can’t be waved away.

You cannot attribute delayed influence. Someone reads your name in a generated answer, closes the tab, and searches your brand three days later on another device. That arrives as direct or branded traffic. Every “AI-influenced pipeline” number you’ll see is a modelling assumption wearing a metric’s clothes.

How to report around the limits

The practical response is to report ranges and publish your error bars beside the number, the way the unattributed-session share sits beside our referral table above. Stating the hole costs nothing and is most of the difference between a measurement programme and reporting theatre. Once the numbers show a real gap, the work moves upstream to the content structures that get extracted and cited in the first place.

For a wider framing of how these layers stack from crawl access through to revenue, Search Engine Land’s five-layer AI search measurement framework is a useful scaffold, and the vendor view of the same problem is set out in Semrush’s AI visibility reporting guide. Read the second one knowing every metric in it maps to a product SKU.

Frequently Asked Questions