If you are measuring AI strategic visibility by counting citations, you are undercounting your actual presence by roughly three times. ChatGPT averages about 2.4 brand mentions per prompt but only 0.74 linked citations, which means most of the times your brand appears in an AI answer, it appears without a link. Citation tracking, which is what most tools report and most dashboards show, therefore captures under a third of the thing you are trying to manage. The metric to build your program on is inclusion rate, not citation count.
This article is about the measurement layer: which metrics exist, how they are calculated, what a defensible baseline looks like, and which numbers in this category not to trust. For the underlying concept, our comparison of GEO vs. SEO covers how these systems differ from search, and for the tooling, our roundup of best AI Overviews rank tracker tools covers what to buy.
The metrics, with their actual formulas
The category has settled on a handful of metrics that are frequently given different names by different vendors. The definitions:
- Inclusion rate. The percentage of tracked AI responses that mention your brand. Formula: (total brand mentions ÷ total tracked responses) × 100. This is the foundational metric and the one to report.
- AI share of voice. Your mentions as a percentage of total mentions across your prompt set, measured against your top three to five competitors. This is the metric that survives a board conversation, because it is relative rather than absolute.
- Citation frequency. How often your domain appears as a cited source. Useful, but per the opening, an incomplete picture.
- Share of model. Total brand mention volume across tracked prompts, expressed as a percentage. Broadly a variant of inclusion rate under a different name.
- Position in AI Overviews. Where you appear in a synthesised answer list, typically first through seventh.
- Brand sentiment. Whether the framing around your mention is positive, neutral, or negative. Undervalued, because being mentioned dismissively is worse than not being mentioned.
- Composite visibility scores. Proprietary weighted formulas. One published example weighs its components 50%, 30%, and 20%. These are fine for tracking your own trend and useless for comparing across tools, because no two vendors weigh the same way.
Reported benchmarks put an inclusion rate of 40% to 60% as strong and below 20% as effectively invisible. Treat those thresholds as directional, since they are set by the companies selling the tracking.
Visibility is now partly decided before retrieval.
This is the structural change that matters most, and it happened recently. Around mid-August 2026, Promptwatch data showed an 86% drop in Reddit citations in ChatGPT responses, correlating with a methodology shift in which ChatGPT now assembles a shortlist of known brands before fetching web results.
Think about what that means for a visibility program. If a shortlist of recognized entities is formed before any retrieval happens, then a well-optimized page cannot get you into an answer you were never a candidate for. The work moves from on-page optimization to entity recognition: being consistently described, in consistent terms, across enough independent sources that you are a known name rather than a retrievable document.
It also means the user-generated-content arbitrage that many brands leaned on has narrowed sharply. Building presence through forum threads was working partly because those sources were cited heavily. An 86% drop in one platform’s citations is a strategy being closed.
Set expectations on traffic honestly.
Before you build a business case, two numbers worth internalizing.
AI engines cite roughly two to seven domains per answer, against about ten results on a traditional search page. Fewer slots, and the slots are filled by a different selection process.
And this one, which should reset anybody’s ROI model: GPTBot has been reported to crawl on the order of 1,500 pages for every single referred visitor click. Whatever AI visibility is worth, it is not worth it as a direct traffic channel at that ratio. The value is in being the recommendation, not in the click.
That reframe matters for how you report. If you present AI visibility as a traffic acquisition program, the numbers will look terrible and the budget will be cut. If you present it as brand presence inside the recommendation layer, measured by share of voice against competitors, the numbers are defensible and the strategy survives.
Content recency is a ranking factor here.
One of the more actionable findings in this area is how strongly these systems favor recent material:
- Around 79% of surfaced content falls in the 2024 to 2025 range, with 89% from 2023 to 2025
- Google’s AI Overviews draw roughly 44% of citations from 2025 content and 30% from 2024
- Perplexity takes about 80% of citations from 2023 to 2025
Two practical consequences. Updating existing high-value pages is probably higher return than publishing new ones, because a refreshed page inherits its existing authority and gains recency. And evergreen content with no visible date is at a structural disadvantage, which is an argument for displaying update dates rather than hiding them.
Separately, about 81% of AI-cited pages use schema markup and roughly 99.3% of citations originate from open-access sources. Neither is a causal claim, but a page behind a paywall or registration wall is effectively invisible to this channel, and that is a decision worth revisiting.
Where the intent is, and where it is not
Mention volume varies enormously by query type, and this is where to concentrate a limited prompt-tracking budget:
- Commercial queries generate roughly four to eight times more brand mentions than informational ones
- Transactional queries generate eight to ten times more
- Prompts containing the word “best” average about 4.8 mentions
- Shopping queries can produce up to 21 mentions in a single answer
Tracking tools charge by prompt volume, so this is a budget decision as much as a strategic one. Fifty commercial and transactional prompts will tell you more about your competitive position than five hundred informational ones. Most programs get this backwards because informational prompts are easier to think of.
A defensible measurement program
- Define a prompt set of 50 to 150 prompts, weighted toward commercial and transactional intent. Write them down and freeze them. A prompt set that changes every month produces a trend line that means nothing.
- Baseline inclusion rate and share of voice before you change anything. Against three to five named competitors. Without a pre-intervention baseline, you cannot attribute any later movement.
- Track mentions and citations as separate metrics. Given the roughly 3x gap between them, collapsing them into one number is how programs report the wrong thing for a year.
- Add sentiment to anything commercial. Appearing in a “best” list in fourth place with a caveat attached is a different outcome from appearing first, and inclusion rate cannot tell them apart.
- Measure across at least three engines. Citation behavior diverges sharply between them, as the Reddit shift demonstrated. A single-engine program is one methodology change away from being wrong.
- Report quarterly, not weekly. These numbers are noisy at short intervals, and weekly reporting produces a reaction to noise.
- Never report AI visibility as a traffic channel. Report it as share of voice in the recommendation layer. See the crawl ratio above.
For the tooling that supports this, our pieces on Ahrefs Brand Radar vs. Scrunch and Ziptie AI search analytics cover the platform options, and AI search optimization services cover the agency route if you are not building this in-house. Entry-level tracking in this category generally runs $29 to $99 a month, which makes a baseline cheap enough that there is no good reason not to have one.
Be skeptical of the statistics in this field.
A closing note that I would want if I were reading this.
Almost every figure circulating about AI visibility comes from a company selling AI visibility tools. Some of it is solid measurement. Some of it is not. The claims I would treat most carefully are the large percentage increases in AI-referred sessions, which usually compare against a near-zero base, and the forward projections about what share of search traffic will migrate by a given year, which are estimates presented as findings.
The figures I have more confidence in are the structural ones: the ratio between mentions and citations, the number of domains cited per answer, the recency distribution, and the crawl-to-click ratio. Those describe how the systems behave rather than predicting what markets will do.
When you build your own program, your own baseline is the only number you should fully trust. Everything above is useful for orientation, and none of it describes your brand.
