There is no validated, industry-standard “AI visibility score” for brands in 2026 — every published mention rate, citation rate, or share-of-voice number is conditional on the engine, sample, date, and query set it was measured against, and the most rigorous critical survey of the field concludes that a single scalar score is defensible only when it’s paired with an explicit objective function, which almost none of the tools selling one publish. This post previews the methodology we’re using to build DigiTortoise’s own Brand AI Readiness measure, and it starts by being honest about what the current research does and doesn’t support.
Why “one score” is harder than the dashboards suggest
The clearest statement of the problem comes from a July 2026 critical survey of 45 generative-engine-optimization studies published since November 2023. It proposes describing a brand or page’s exposure as a visibility vector rather than a single number: discoverability, context exposure, citation probability, observable prominence, absorption into the generated answer, factual fidelity, and downstream behavioral outcome. Its conclusion on aggregation is blunt: “a scalar score is defensible only when the weights correspond to an explicit objective. Aggregating a mention, an accurate citation, and a conversion without a utility model merely obscures normative choices” (Martinez, arXiv:2607.14035, 15 Jul 2026).
That’s a direct challenge to every product — including ones we’ve covered in our own AEO tools landscape — that outputs a single “AI visibility score” without disclosing how the sub-metrics are weighted. It doesn’t mean the number is useless. It means it needs a denominator and a definition attached, every time.
The nine-level hierarchy: what each metric actually establishes
The same survey organizes the measurable stages of “getting cited” into nine levels, each with an explicit note on what it does and doesn’t prove. We’re adopting this hierarchy as the backbone of our own methodology rather than inventing a new taxonomy:
| Level | What it measures | What it does not establish |
|---|---|---|
| Activation | Whether the engine triggers a web search / retrieval step at all | Whether your content is found once search runs |
| Retrieval | Whether your page enters the candidate document set | Whether it survives reranking into the model’s context |
| Mention | Whether your brand name appears anywhere in the answer | Whether it’s cited, sourced, or framed positively |
| Citation | Whether a clickable source link points to your page | Whether the citation accurately reflects your content |
| Prominence | Position/rank of the mention or citation in the answer | Whether position actually predicts clicks (rarely tested) |
| Coverage | Word count or share of the answer attributed to your source | Reader attention or comprehension |
| Absorption | How much of the answer’s facts/language trace to your content | Whether the trace is faithful to your original claim |
| Fidelity | Whether the citation accurately supports the claim it’s attached to | Business outcome |
| Behavior | Clicks, referral traffic, conversions attributable to the mention | Almost never measured in published studies |
Most vendor “visibility scores” collapse straight from Mention to a headline percentage, skipping Citation, Fidelity, and Behavior entirely. That’s the gap our methodology is built to be explicit about rather than paper over.
What the 2026 production data actually shows
The largest dataset we found tracking this at scale comes from a June 2026 study covering 102 brands, 3,508 tracking runs, and 102,025 prompt responses across five engines (ChatGPT, Gemini, Perplexity, Claude, Grok) between March and May 2026. Its headline finding is a “brand-stature visibility ladder”: already-prominent brands (n=11) were mentioned in 72.9% of unbranded day-1 prompts (95% CI 60.1–84.2%), mid-tier brands (n=36) in 43.6% (36.4–50.9%), and smaller brands (n=55) in just 11.4% (4.2–20.3%) — roughly a 30-percentage-point drop per stature tier (Kruskal-Wallis H=38.32, p=4.78×10⁻⁹) (Kumar, arXiv:2606.20065, 18 Jun 2026). The authors are explicit that this is observational, not causal: brand stature wasn’t randomly assigned.
The same dataset found that source-link inclusion varies enormously by engine: in a 2,500-response CRM reference study, Perplexity attached a clickable source 95% of the time, Gemini 35%, Grok 20%, ChatGPT 15%, and Claude just 10%. And sentiment is far less stable than presence: “whether the AI frames a brand positively or negatively flips about 6.7 times more often than whether the brand is mentioned at all.” That instability is one reason our methodology treats sentiment as a separate, lower-confidence layer rather than folding it into a single composite score.
It’s also why our 7-layer AI visibility audit and this methodology track citation and mention rates separately by engine rather than blending them into one cross-platform average up front.
Where the “GEO boosts visibility 40%” claim actually comes from — and why it doesn’t generalize
Almost every AEO/GEO pitch deck cites some version of “generative engine optimization increases visibility by up to 40%.” That number traces to a specific 2024 KDD paper that introduced Position-Adjusted Word Count (PAWC) and an LLM-judged Subjective Impression metric, tested against a fixed five-document context window. The best method (quotation addition) moved PAWC from a baseline of 19.3 to 27.2 — a 41% relative gain, on that one metric, in that one simulated setup (Aggarwal et al., arXiv:2311.09735, KDD ’24). It also found that citing sources helped low-ranked documents dramatically (+115.1% for the 5th-ranked source) while hurting the top-ranked one (-30.3%), and that keyword stuffing did nothing or made things worse.
The 2026 survey is pointed about the reuse of that number: “the figure is a relative maximum on one metric under a specific configuration,” not a general promise about ranking in ChatGPT (Martinez, arXiv:2607.14035). It also notes a real methodological risk in the original paper: the same model family that generated the rewrites also judged them, a circularity concern the survey flags as needing separated generator/judge models and human validation going forward.
What this means for our methodology
Given that, our Brand AI Readiness preview will report each layer of the hierarchy separately, by engine, with the sample size and date window attached — not a single blended score. We’ll track: mention rate and citation rate (kept distinct, since they diverge by 10–85 percentage points depending on engine); position/prominence where an engine exposes it; sentiment as a volatile secondary signal; and structural readiness factors we already document independently, including how AI assistants decide which brands to recommend, entity resolution via structured identity data, and referral attribution via our guide to measuring AI referral traffic. Where we can’t establish fidelity or downstream behavior — the two levels the literature measures least — we’ll say so rather than imply a number we don’t have.
FAQ
Is there an industry-standard AI visibility score?
No. Multiple commercial tools publish a composite 0–100 “AI visibility score,” but none of the underlying weighting methodologies has been independently validated, and the most recent academic survey of the field argues a scalar score is only defensible when its weights correspond to a stated objective — which vendor dashboards typically don’t disclose.
What’s the difference between a “mention” and a “citation”?
A mention is your brand name appearing anywhere in an AI-generated answer. A citation is a clickable source link pointing to your page. The gap between the two is large and engine-dependent: one 2026 study found source-link inclusion ranging from 95% (Perplexity) down to 10% (Claude) even when the brand was mentioned.
Why do published AI Overview activation rates vary so much between studies (13.7% vs. 51.5%)?
Because activation rate — whether an engine triggers web search/retrieval at all — depends heavily on query phrasing, sample composition, and date. One 2026 study of 55,393 trending queries found a 13.7% overall AI Overviews activation rate that rose to 64.7% for question-phrased queries; a separate 11,500-query study found AI Overviews shown 51.5% of the time. Neither number is wrong; they describe different samples.
Last updated September 7, 2026. This page is refreshed as benchmarks and scores move.
Pingback: How AI Assistants Decide Which Brands to Recommend - Tech Blog