TL;DR — A 2026 study put 280 B2B software buyer questions to ChatGPT, Claude, Perplexity, and Google AI Overviews and found that 63.3% of vendor mentions came from a single engine — only 17.5% appeared on all three. Being “on ChatGPT” is the modal outcome, not the goal. If your AEO dashboard only probes one or two engines, you are flying blind on the engine your actual buyer used.
The study (and its honest caveats)
Broadcastwell’s 2026 State of GEO (data collected 4 Aug 2026, posted to SSRN 10 Aug) ran 280 buyer questions across four production engines. It’s an industry study, not peer-reviewed, and it was honest about its own limits: three of four API accounts ran out of credit mid-collection, so the four-engine intersection is 18 questions, not 280. The operator also sells in the category it measures and excluded itself from rankings.
The verified numbers:
- 63.3% of 2,238 distinct vendor mentions across 142 questions answered by Claude + Perplexity + Google AIO came from one engine only.
- 17.5% of mentions appeared on all three engines.
- Mean pairwise Jaccard similarity of vendor sets: 0.313 (95% CI 0.295–0.330) — most engines agree on very little.
- Google AIO returned an AI Overview for 92.1% of buyer questions.
- Google AIO’s within-engine agreement gap (+0.258) survives a first-five-vendors length control.
- Claude’s raw gap (+0.165) vanishes under the same length control (+0.002, CI includes zero) — a pooled “AI search rank” without a length control is suspect.
What this means for your dashboard
Yesterday’s Cloudflare AEO dashboard — like most off-the-shelf AEO panels — still probes only Claude and GPT. It cannot see Perplexity or Google AIO. If being named on a single engine is the modal outcome, then “we showed up in ChatGPT” or a Cloudflare Mention Rate tells you almost nothing about the buyer who used Perplexity.
Treat any single-engine dashboard as a crawl/referral panel, not a scoreboard.
What to do instead
- Run the same buyer prompts on all four engines — category questions, comparison questions, and “best X” questions, exactly as your buyer would type them.
- Log named vs cited per engine. Being named is presence; being cited is source. They are different games.
- Never average engines. A blended “AI visibility score” hides the exact failure this study found.
- Close the gap on the engine you’re missing, not the one you’re winning.
The extractability side
This is partly a measurement problem and partly a surface problem. The pages engines can actually read and quote — clean markdown, sequential headings, fresh content, llms.txt indexes — are the same surfaces that make you citable on any engine.
At VijayaTech Labs, we built this directly into our publishing architecture:
- We cover the llms.txt v2 pieces (markdown alternates +
rel="describedby") in the Glint AEO guide. - We align this with structured third-party corroboration across AI-readable launch directories to maximize multi-engine consensus.
- We run autonomous multi-engine telemetry inside our Hermes Agent ops stack.
Frequently Asked Questions
Why do different AI engines recommend completely different vendors?
Each answer engine relies on distinct retrieval architectures, training cutoffs, and real-time grounding sources. Perplexity indexes real-time web pages, Google AI Overviews relies on the Google Search Index, while ChatGPT and Claude blend parametric memory with specific search integrations. This results in an average vendor overlap (Jaccard similarity) of only 31.3%.
What is the difference between being named vs cited in an AI answer?
Being named means the LLM mentioned your company or product in its generated text. Being cited means the engine included a direct clickable hyperlink to your domain as an authoritative source. Citable content requires clean, crawlable markdown surfaces (like llms.txt and raw markdown twins).
Sources
- Broadcastwell / Sivakumar, Divergence Survives a Length Control… (4 Aug 2026, SSRN 10 Aug) — https://ssrn.com/abstract=7234418 · DOI 10.5281/zenodo.21789120 · open data: https://github.com/Broadcastwell/state-of-geo-2026
- Cloudflare AEO dashboard docs (probes Claude + GPT only)