Manifesto · September 2026
The models have already decided what every product is
Ask ChatGPT, Claude, or Gemini to recommend software and the answer comes out of a memory, not a market. Somewhere in that memory, each brand sits in a bucket: the CRM, the ticketing tool, the drag-and-drop app builder. The bucket decides which shortlists the brand makes, which caveats follow its name, and which rival takes the deal. Most companies have never seen theirs.
Label lag: the models describe companies as what they used to be
Thunderdome runs scripted buyer conversations against the three major models at scale and reads back what they believe. 7 companies have been through the full battery so far. Not one is described the way its own homepage describes it.
- Close 98% of runs file it as SMB sales CRM with a built-in dialer. The homepage says AI-powered sales CRM with an autonomous AI teammate. 1% of runs agree.
- Linear 98% of runs file it as sleek issue tracker for startup engineering teams. The homepage says system for modern product development. 0% of runs agree.
- Retool 97% of runs file it as drag-and-drop / low-code builder. The homepage says AI app generation platform. 3% of runs agree.
- Klaviyo 90% of runs file it as marketing automation platform. The homepage says B2C CRM built on customer data. 2% of runs agree.
- Zendesk 72% of runs file it as omnichannel customer experience suite. The homepage says AI-powered customer service platform. 11% of runs agree.
- Braintrust 70% of runs file it as prompt-testing and evals tool. The homepage says enterprise platform for building reliable AI products. 1% of runs agree.
- Lattice 47% of runs file it as talent management suite. The homepage says all-in-one people platform (HRIS, payroll, performance). 18% of runs agree.
Retool shows the mechanism most completely. Its homepage last led with drag-and-drop language in 2023, verified against archived copies of the site. Three years of repositioning later, 97% of aided runs still file it there. Model memory moves slower than positioning. That gap is label lag, and it compounds quietly: a mislabeled brand is not rejected, it is simply never surfaced when the right buyer asks the right question.
Visibility metrics count appearances. Buying decisions follow beliefs.
The emerging playbook for AI search borrows its instruments from SEO: count mentions, track rank, chase citations. Those numbers answer one question, whether the model says the name. They cannot answer the three questions that decide a recommendation: as what, with which caveats, on whose evidence.
The measured record shows why the distinction matters. Zendesk appears in nearly every grounded answer its battery collects, about as visible as a brand can be. It also drew a hedged verdict in 120 runs out of 120, with the models parking it between the legacy pole it left and the AI pole it claims. A brand can rank first in AI visibility and still be filed under the wrong decade. Visibility measures presence. The problem is identity.
Perception lives on three surfaces, and they rarely agree
What a market believes about a product is legible in three places. The claim: what the company's own site says, page by page. The memory: what the models say unprompted, from training. The record: the third-party pages the models actually read and cite when they search. Perception is the degree to which the three agree, and every disagreement has a type and a fix.
Relevance gap
The company claims it; the models never echo it.
Retool describes multi-environment deployment on 17 pages of its site. The grounded answers almost never credit it.
Rebuttal gap
Buyers raise an objection; the site never answers it.
The models call Retool’s per-user pricing expensive at scale in 9 separate assertions. No page on the site addresses the objection, so reviews and rivals fill the silence.
Free equity
The models grant a strength the site barely claims.
The models already call Retool the industry standard for internal tools and praise its built-in managed Postgres. Claiming a belief the models already hold is the cheapest positioning win available.
Contradiction
The claim and the record disagree, so the models hedge.
Retool’s site claims non-technical teams can build; the models keep saying it requires SQL and JavaScript. Contradictions convert clean yeses into walls of maybes.
Examples above are from the Retool triage: an 80-page crawl of the commercial site, decomposed into claims, cross-referenced against 4,289 assertions harvested from grounded model answers.
Naming the category: perception analytics
Perception analytics is the discipline of measuring what AI models believe a product is, why they hedge on it, and which sources taught them, then closing the gap between the claim and the belief. It is to AI visibility what brand tracking is to share of voice: the layer underneath, the one that explains the number instead of restating it.
The discipline has five requirements, and anything missing one is a mention counter:
- It reads beliefs, not rankings: descriptor buckets, verdicts, and reason codes taken from the models' own words.
- It forces trade-offs: head-to-head choices at scale, because open-ended praise hides the real ranking.
- It traces provenance: grounded answers with the domains they cite, the record the models actually read.
- It triangulates all three surfaces, claim against memory against record, so every gap routes to a specific fix.
- It repeats on a clock: perception is a time series, and a fix is only proven when the belief moves.
Put as a from-to: from counting the mentions in AI answers, to auditing the beliefs behind them.
The instrument is live and free
The Perception Lab is the working instrument of perception analytics, and every lab on this site is public: no signup, no gating. Each one publishes the full battery for a single company, the verdicts, the forced choices, the reason codes, the grounding record, and the search-demand backdrop, narrated end to end. The index that surrounds it tracks visibility across the B2B SaaS market, because presence still matters. The lab explains what presence cannot.