An AI visibility score is not a measurement. It is an average of at least four different measurements — mentions, citations, rankings and outcomes — that now move in opposite directions, so the composite can rise while every number a business cares about falls. The fix is not a better score. It is refusing to blend the layers in the first place, and naming which one moved before you explain why.
Demand for these scores is real and growing fast: Ahrefs Keywords Explorer data shows searches for "ai search tracking" up 184% and "ai rank tracking" up 175% year over year. What has not kept pace is agreement on what any of it counts.
Key takeaways
- Across 863,000 keyword SERPs and 4 million AI Overview URLs, only 37.1% of cited pages also held an organic top-10 position — down from roughly 76% in the same study's July 2025 run.
- ChatGPT cited just 49.98% of the URLs it retrieved across 1.4 million prompts. Retrieval and citation are separate events, and most tools can only see the second.
- Citation rate depends overwhelmingly on where a page was retrieved from: 88.46% for general search results versus 1.93% for Reddit and 0.40% for academic sources.
- Adding JSON-LD to 1,885 pages produced +2.4% in AI Mode and +2.2% in ChatGPT — neither statistically significant — while AI Overview citations fell 4.6%.
- With an AI Overview present, position 1 now takes a 58% lower clickthrough rate, up from 34.5% in the prior study. A score measuring presence cannot see that.
What is an AI visibility score actually measuring?
Usually several things at once, averaged into a single trend line:
| Layer | The event being counted | What it tells you |
|---|---|---|
| Mention | Your brand name appears in an answer | Whether the model associates you with the topic |
| Citation | A URL of yours is linked as a source | Whether a specific page earned attribution |
| Retrieval | Your page was fetched but may not be shown | Whether you are in the consideration set at all |
| Ranking | Your position in classic organic results | Whether traditional search still feeds the answer |
| Outcome | Sessions, conversions, revenue | Whether any of the above paid |
A mention can happen with no link attached. A citation can happen with no brand name in the text. A page can be retrieved and never surface. These are not degrees of the same thing — they are different events with different causes, and a tool that adds them together produces a number whose movement is uninterpretable by construction.
Does ranking in Google's top 10 still get you cited?
Less than half the time, and the exact answer depends on a methodological choice most write-ups do not disclose.
Analysing 863,000 keyword SERPs and 4 million AI Overview URLs, Ahrefs found this distribution:
| Cited URL's organic position | Counting all SERP blocks | Counting organic links only |
|---|---|---|
| Top 10 | 37.9% | 37.1% |
| Positions 11–100 | 31.2% | 26.2% |
| Not in top 100 | 31.0% | 36.7% |
Both cuts come from the same dataset. Whether you say "31% of AI Overview citations come from pages that do not rank at all" or "36.7%" depends entirely on whether featured snippets, People Also Ask boxes and video packs are counted as ranking. That is a five-point swing produced by a definition, not by Google.
This is the argument applied to its own evidence. If a single well-documented study yields two defensible numbers for the same question, a proprietary score aggregating dozens of undocumented signals is not going to yield one.
The direction, though, is unambiguous. The same study run in July 2025 put top-10 pages at roughly 76% of citations. It is now roughly 38%. Google's query fan-out — decomposing a question into sub-queries and drawing sources across the whole cluster — means pages that rank for nothing you tracked can be cited for something you did not think to track.
Why does the same page get ignored in one system and cited in another?
Because citation rate is largely a property of the retrieval channel, not the page. Across 1.4 million ChatGPT prompts, the split by source type was severe:
| Source type | Citation rate once retrieved |
|---|---|
| General search results | 88.46% |
| News | 12.01% |
| 1.93% | |
| YouTube | 0.51% |
| Academic | 0.40% |
Overall, ChatGPT cited 49.98% of retrieved URLs — 23.4 million of them — and left the rest invisible. 67.8% of all non-cited URLs came from Reddit alone.
That last figure is the one worth sitting with. A brand thoroughly discussed on Reddit is being read constantly and credited almost never. It is shaping the answer while registering as zero on every citation-based tracker. Conversely, a page reaching the model through general web search starts from an 88% citation rate before anything about its content is considered.
Two page-level signals did correlate meaningfully in that dataset: cited pages' titles matched the model's internal sub-queries more closely (0.656 cosine similarity), and results with natural-language URL slugs were cited 89.78% of the time versus 81.11% without. Both are cheap to act on and neither is what most AI visibility dashboards report.
Does schema markup increase AI citations?
On the available causal evidence, no — despite a correlation strong enough to have launched a small industry.
Across 6 million URLs, pages cited by AI were almost three times more likely to carry JSON-LD, with 53% of AI-cited pages having schema. That correlation is real. It is also exactly what you would expect if structured data simply tends to live on better-maintained sites that publish stronger content and earn more links.
Ahrefs then tested it directly: 1,885 pages that added JSON-LD between August 2025 and March 2026, matched against 4,000 control pages with similar prior citation levels.
| Platform | Change after adding JSON-LD | Statistically significant? |
|---|---|---|
| Google AI Mode | +2.4% | No |
| ChatGPT | +2.2% | No |
| Google AI Overviews | −4.6% | Yes |
No meaningful uplift anywhere, and the one significant result went the wrong way. Structured data still earns rich results and still helps machines parse entities — there are good reasons to ship it. "It will get us cited by AI" is not currently one of them, and any visibility score that rose after a schema rollout was almost certainly measuring something else.
Can Google Search Console close the measurement gap?
This is the honest complication in every "just measure outcomes instead" recommendation, and it is usually skipped.
Google's Generative AI performance report, rolled out to all sites by 31 August 2026, covers AI Overviews and AI Mode. It reports impressions only — no clicks, no CTR, no average position, no query data. You can see that a URL appeared in an AI surface. You cannot see whether anyone clicked it, or what they asked to get there.
So the outcome layer that is supposed to discipline the visibility layer is, inside Google's own first-party tooling, not yet an outcome layer. It is a second impression count.
That matters because of what happened to clicks in the meantime. Comparing December 2023 with December 2025 across 300,000 keywords using aggregated Search Console data, the presence of an AI Overview correlated with a 58% lower CTR for the top-ranking page — up from 34.5% in the earlier version of the study. And the damage is not evenly distributed:
| Organic position | CTR change with an AI Overview present |
|---|---|
| 1 | −58.0% |
| 2 | −50.8% |
| 3 | −46.4% |
| 5 | −32.6% |
| 10 | −19.4% |
The pages losing most are the ones a ranking-weighted visibility score rewards most. A brand can hold position 1, gain AI citations, watch its composite score climb, and lose more than half its clicks — with every input to that score behaving exactly as designed. This is the same blind spot that makes AI Overview layout changes impossible to evaluate from Google's data alone.
What should you measure instead?
Four things tracked separately, never averaged:
| Metric | Question it answers | Where it comes from |
|---|---|---|
| Mention rate | Does the model associate us with this topic? | Prompt sets run at volume, not single queries |
| Citation rate | Are our URLs earning attribution? | Per-platform, never blended across engines |
| Ranking | Is classic search still feeding the answer? | Standard rank tracking |
| Outcome | Did any of it produce revenue? | Analytics and referral data, not Search Console alone |
Three practical rules follow.
Run prompt sets, not prompts. These systems are probabilistic — the same question produces different answers across runs. A single prompt checked weekly measures variance. Only volume produces a rate.
Never blend platforms. A change in ChatGPT and a change in AI Overviews have different causes and different fixes. Averaging them destroys the diagnostic value of both.
Diagnose before reacting. When a number moves, the useful question is which one:
| What moved | Likely cause | Where to look first |
|---|---|---|
| Citations fell, rankings flat | Retrieval or fan-out coverage changed | Compare cited vs. uncited pages for title and sub-query alignment |
| Mentions fell, citations flat | A competitor gained topical association | Prompt-set share against named rivals |
| Score rose, clicks fell | Presence increased, link value did not | CTR by position against the AI Overview ladder above |
| Everything moved at once | Measurement changed, not reality | Verify the tool's methodology before writing the report |
That last row is the most common and the least investigated. Before explaining a score change to anyone, confirm the score is still counting the same things it counted last month.
Frequently asked questions
Is a brand mention the same as an AI citation?
No. A mention is your brand name appearing in generated text, usually with no link. A citation is a URL of yours listed as a source, which can happen without your brand being named at all. They have different causes and should be tracked as separate metrics.
Why do two AI visibility tools report different scores for my brand?
Because they count different events and weight them differently. Even a single well-documented study can produce two defensible answers to the same question: Ahrefs' AI Overview research reports that either 31.0% or 36.7% of cited URLs fall outside the organic top 100, depending purely on whether non-organic SERP blocks count as ranking.
Does adding schema markup get my pages cited by AI?
Current evidence says no. Tracking 1,885 pages that added JSON-LD against 4,000 matched controls, Ahrefs measured +2.4% in AI Mode and +2.2% in ChatGPT — neither statistically significant — and a statistically significant 4.6% decline in AI Overviews. Schema still earns rich results; it is not a citation lever.
Does ranking in Google's top 10 still matter for AI Overviews?
It helps but no longer dominates. Only 37.1% of AI Overview citations came from pages in the organic top 10, down from roughly 76% in the study's July 2025 run. Query fan-out means pages ranking for related sub-queries can be cited without ranking for the original one.
Can I measure AI search performance in Google Search Console?
Only partially. The Generative AI performance report covers AI Overviews and AI Mode but reports impressions only — no clicks, CTR, average position or query data. It confirms you appeared; it cannot tell you whether appearing was worth anything.
How often should I run AI visibility checks?
Frequently enough to build a rate rather than a snapshot. Because responses are probabilistic, single prompts checked occasionally measure randomness. Run a fixed prompt set at consistent volume and treat the aggregate percentage, not any individual answer, as the signal.
Where to start
The cheapest first move is confirming the AI systems can reach your pages at all — citation analysis is meaningless for content that was never retrievable. Run your domain through the AI crawler checker to see which agents are permitted, and the meta tags checker to confirm titles are doing the sub-query matching work the ChatGPT data rewards.
From there, defining the AI search KPIs you will actually report on matters more than which tool you buy. For the broader picture of why presence and clicks have decoupled, see why clicks are falling while brand demand isn't.
Sources
- Ahrefs — 5 AI Search Trends in 2026
- Ahrefs — AI Overview citations and top-10 rankings
- Ahrefs — Why ChatGPT cites one page over another (1.4M prompts)
- Ahrefs — We tracked 1,885 pages adding schema
- Ahrefs — Update: AI Overviews reduce clicks by 58%
- Google Search Console Help — Generative AI performance report