State of AI Search 2026: The Measurement Gap
In late August 2026, 171 search professionals from 36 countries answered SEOFOMO's State of AI Search Optimization survey, run by Aleyda Solis between August 17 and 27. The headline movement is adoption: budgets, strategies, and tracking all jumped hard in one year. The report's own conclusion is blunter, and more interesting: the industry has stopped arguing about whether to do AI search optimization and started doing it, but it still can't measure it properly.
That measurement gap has a specific shape, and the survey's numbers reveal it. This post walks through what the report found and why the most common measurement setups can't answer the question being asked of them. Then it shows what our own measurement across a portfolio of sites reveals about the split a single visibility number hides.
Quick Navigation
- What the Survey Found
- The Measurement Gap in the Numbers
- Ranking and Citation Are Different Instruments
- What Our Own Measurement Shows
- How to Start Measuring the Split
- Frequently Asked Questions
What the Survey Found
The 2026 SEOFOMO survey shows AI search optimization crossing from debate to standard practice in a single year: budget allocation nearly doubled to 69%, dedicated strategies jumped from 51% to 75%, and tracking reached 92% of respondents. The adoption question is settled.
| Measure | 2025 | 2026 |
|---|---|---|
| Have budget allocated for AI search | 38% | 69% |
| Dedicated AI search strategy (at least some sites) | 51% | 75% |
| Tracking AI visibility and citations | 78% | 92% |
| Asked about AI visibility by a client or decision maker | 91% | 98% |
| Nobody owns the work | 11% | 4% |
Ownership settled too: 77% of respondents say SEO teams lead the work, essentially unchanged from 2025, which quietly answers the "will GEO be a separate discipline" debate for now. And 61% say investment grew during 2026 against 1% who saw it cut. Whatever else is uncertain, the direction of spend isn't. All figures are from the SEOFOMO State of AI Search Optimization report, August 2026.
The Measurement Gap in the Numbers
The same survey that shows 92% of teams tracking AI visibility also shows 28% observing no measurable impact and 35% unable to quantify what AI search contributes to revenue, which means a large share of that tracking produces numbers nobody can act on. Unreliable measurement leads the challenge list at 19%. Another 14% can't tell what drives visibility, and 14% more can't connect it to conversions.
The tools list explains a lot. The most-used measurement tools for AI search were Semrush at 29%, Google Search Console at 28%, and Ahrefs at 24%. Those are excellent instruments, built over two decades to answer one question: where does this page appear in a ranked list? AI search answers ask a different question entirely, and pointing a rank tracker at a citation problem produces exactly the survey's result: lots of tracking, little understanding.
The metric fragmentation shows the same confusion from another angle. Respondents split across at least six different quantities:
- Brand visibility and mentions (19%)
- Traffic from AI platforms (16%)
- Inclusion in cited sources (15%)
- Competitive share of voice (14%)
- Brand links inside answers (12%)
- Sentiment (11%)
No metric reached a fifth of respondents, because the industry hasn't yet agreed on what the underlying quantity even is.
Ranking and Citation Are Different Instruments
Ranking measures where your page appears when a system sorts results; citation measures whether a model quotes your page when it composes an answer, and the second depends on mechanics the first never touches: whether the model retrieves your page at all, and whether your content is structured so an answer can be lifted from it cleanly. A page can rank in the top three and never earn a single AI citation.
The extraction side is the best-researched. A 2025 study by Zyphra found that 72.4% of blog posts cited by ChatGPT share one structural trait, the answer capsule: a self-contained, quotable statement placed directly after a heading. Princeton's GEO research measured visibility gains of up to 41% from adding citations and quotable statistics to pages. None of that is rank tracking. It's a different discipline with its own research base, and it's why "we rank well" and "we get cited" diverge so often in practice.
But extraction is only half. The other half is retrieval, and that's where the most useful measurement split lives.
What Our Own Measurement Shows
Every month we measure our own portfolio of sites the same way: buyer-phrased prompts run against ChatGPT and Claude through their APIs, each prompt executed twice, once with web search enabled and once without. In August 2026 that was about 2,240 evaluations, and the with-search versus without-search split diagnosed two opposite problems that a single mention-rate would have blended into one number.
The without-search run measures memory: does the model know the brand from training? The with-search run measures retrieval and selection: when the model actually looks at the live web, does it find you and choose to cite you? Reading them as a pair is the diagnostic:
- Our established site was mentioned in 12.7% of memory runs and 14.7% of grounded runs. Searching barely helps, a gap of two points. The models find the site and cite a bigger name anyway. That is a competitive-selection ceiling, and publishing more of the same content will not move it; being the more extractable answer on specific prompts might.
- Our youngest revenue site was mentioned in 0% of memory runs and 7.2% of grounded runs. The models have never heard of it, but when they search, they find it. That is the healthy early-stage pattern: visibility that time and publishing volume genuinely do fix.
- The engines disagree sharply. On that young site, Claude's search surfaced it at 12.5% while ChatGPT's surfaced it at 1.9%, roughly a six-fold spread on identical prompts. A team tracking only one engine would draw the wrong conclusion about the whole channel.
The same 7% grounded rate means opposite things depending on the memory rate next to it. Zero-and-seven says keep publishing. Thirteen-and-fifteen says stop publishing more and start winning specific answers. That distinction is invisible in every single-number visibility score, and it changes what you do next month, which is the entire point of measuring.
How to Start Measuring the Split
You can run a small version of the memory-versus-retrieval split without buying anything: ten buyer-phrased prompts, each asked twice in a fresh chat, once with web browsing enabled and once with it off, recording whether your brand is mentioned and whether it's linked. Twenty runs per engine, repeated monthly, beats most dashboard visibility scores because the pair tells you which problem you have.
Read the results the way the pattern table above reads: grounded-only mentions mean you're findable but not yet famous, and volume plus consistency is the play. Mentions in neither lane on prompts you should win means a retrieval or extraction problem, and restructuring pages so the answer sits high and quotable is the play. The Brass-SEO AI Citability button covers that structural half for any page you enter: it analyzes ten citability factors, answer capsules included, and returns ranked priority fixes with rewrite suggestions. And since the tools you already pay for mostly measure ranking, knowing which question each instrument answers is the cheapest upgrade available.
The SEOFOMO survey says the industry's next problem is measurement. The split is where we'd start.
Frequently Asked Questions
What did the 2026 SEOFOMO State of AI Search survey find?
The survey of 171 search professionals from 36 countries, run August 17–27, 2026, found adoption surging: 69% now have budget for AI search optimization (up from 38% in 2025), 75% have a dedicated strategy, and 92% track AI visibility. The gap it identified is measurement: 28% see no measurable impact, 35% can't quantify revenue contribution, and the top-cited challenges are unreliable measurement and connecting visibility to outcomes.
Why is AI search visibility hard to measure?
Because the industry is mostly measuring it with ranking-era instruments. The survey's most-used tools for AI search were Semrush, Google Search Console, and Ahrefs, which answer where a page appears in ranked results. AI citation depends on different mechanics: whether a model retrieves the page and whether an answer can be extracted from it cleanly. Tracking also fragments across at least six different metrics, so teams collect numbers without agreeing on what quantity they represent.
What is the difference between ranking and being cited by AI?
Ranking is position in a sorted list of results. Citation is a model choosing your page as source material while composing an answer, which requires the model to retrieve your page and find a liftable, self-contained statement in it. Research by Zyphra found 72.4% of ChatGPT-cited posts contain such answer capsules, and Princeton's GEO work measured visibility gains up to 41% from quotable, cited content. A page can hold position one and still never be quoted.
How do you measure AI search visibility yourself?
Run the same buyer-phrased prompts through an AI engine twice, once with web search on and once off, and record brand mentions in each lane. The pair is the diagnostic: mentions only when search is on means the brand is findable but not yet known, which volume fixes. Mentions in neither lane on prompts you should win points to a retrieval or extraction problem, which page structure fixes. Repeat monthly across more than one engine, because engines can differ several-fold on the same prompts.
Measure the Question You're Actually Asking
The industry spent 2026 proving it will fund AI search work. The next year decides who can tell whether the work is working. Brass-SEO reads your real Google data, and the Brass-SEO AI Citability button analyzes any page's chances of being cited by AI systems, with ranked fixes. Start the three-day free trial and check the pages you most need cited. The full guide to the discipline is at Generative Engine Optimization.