Seven AI engines now answer your buyers before they click a single link. We measure whether those engines name you for the questions buyers actually ask, then hand your team the fixes. No promised numbers. Just measurement you can re-run.
We measure whether the AI engines name a business when buyers ask the category question. Nine categories, three runs each, seven AI surfaces. One result keeps holding, and it is the one the AEO industry would rather you did not see.
The engines that answer from training data instead of searching the live web barely cite local businesses' own websites at all. Claude, on the same questions where web-searching engines reach 38 to 66 percent, collapses to near zero.
| Category | Top web engine | Claude |
|---|---|---|
| Honolulu HVAC | ChatGPT search, 66% | 2% |
| Honolulu med spas | Gemini, 64% | 2% |
| Hawaii CPA firms | OpenAI, 60% | 1% |
| Austin CPA firms (cross-geo control) | holds outside Hawaii | 0% |
The HVAC line is the one we pre-registered. Before a single question ran, we committed the prediction to a public file with a timestamp. We measured two percent. The same story repeats in med spas and accounting, across three unrelated industries in two states. A business can look fine on one surface and be invisible on another. Most vendors cannot publish a table like this, because to publish it you have to actually measure, name the date, and lock the questions before you look. A vibe cannot be put in a table.
Every number here is on the record, with the run and the open-source code that prove it →
Hawaii Theatre Center, the 1922 landmark in downtown Honolulu, agreed to be named. We measured them against a 19-question buyer set, handed their team a forensic memo of where the AI engines named them and where they did not, and they shipped the fixes. We never touched their site.
The execution was theirs. The measurement, before and after, was ours. Because we never touch the property, there was nothing we could have done to flatter the number.
A Google search gives you ten links. An AI answer gives you three names. You can check a Google rank any time. You cannot see an AI answer that left you out.
NeverRanked measures what the AI answer engines cite for your category, split across two layers that fail in different ways.
The deliverable is a forensic research memo and a prepped punch list, ordered by impact. Your team executes it. We do not. That separation is structural.
Atlas is the data-interpretation layer of your dashboard. It answers what the measurement shows. It refuses to tell you what to do. That separation is the engagement.
The boundary is structural. Prioritization lives in your monthly memo, written by the principal. Atlas holds the data. Crossing that line would damage the engagement.
This one is built to be. Four reasons the numbers can be trusted.
Every methodology claim is anchored in a hash-locked pre-registration before the test runs. The claim cannot move after the data lands.
The measurement and aggregator code is on GitHub. One of the seven engines, Gemma, is open-weight, so your own analyst can re-run the prompts and reproduce the numbers without us in the room.
Seven surfaces measured every day. An AI answer can change between two askings of the same question, so a single snapshot is not measurement.
We never touch your website, your code, or your CRM. That is not only a security posture, it is what makes the number trustworthy. The moment the measurer is also the one being measured, the score stops being a measurement and becomes a sales document. We keep our hands off the property on purpose, so the only thing we can do is report what the engines actually cite.
Five stages. Plain words. No SaaS dashboard between you and the work.
Lock the category, the cohort, and the 18 buyer questions we will measure. One call, no homework.
See an example question set →We measure daily across 7 AI tools. Same questions, same hash, every run apples to apples.
PDF or markdown to your team. Named competitors, observed gaps, the clear list of what to fix first.
You ship the work. We measure whether it lands. That separation is the whole position.
What moved, what did not. Updated punch list. Drift alerts when a competitor moves in your category.
The first research memo arrives three weeks after the scoping call.
Every teardown is built from a hash-locked question set, 3 measurement runs, and the same 7 AI tools. Anonymized at the firm level for non-customers, named in full inside paid engagements. The newest, Honolulu HVAC, is the first finding we pre-registered: the prediction was committed to a public timestamp before the measurement ran. The Claude training-data collapse now holds across three unrelated industries.
Before running a single query, we committed a prediction to a public timestamp: Claude would cite Honolulu AC companies under 5% of the time. It came in at 2%, while the web-searching engines reached 38% to 66% on the same questions. A forecast made before the data, not a case study written after it. This is the third unrelated industry where Claude collapses on local firms, after CPA and med spas. Read the pre-registered teardown →
21-bank cohort. The widest cross-engine gap of any category measured. One bank owns the head queries on training-data tools. The long tail sits open on Copilot.
42-firm cohort. AI defers to lead-gen aggregators (SmartAsset, Unbiased, Plannersearch) more than to any individual firm. The structural ground for firms sits outside that middleman tier.
46-practice cohort. Microsoft Copilot cites zero practice websites for the entire cohort. The first Honolulu practice that ranks first on Bing organic effectively owns the Copilot answer.
33-firm cohort. The dominant firm gets roughly three times the citations of the second-tier firms. For any firm outside the top 5, the closable ground is the long tail and Microsoft Copilot.
41-firm cohort. Claude and Gemma cite Hawaii CPA firms less than 2% of the time. Competitive game plays inside OpenAI, Gemini, Perplexity, and Google AI Overviews.
37-firm Austin cohort. First non-Hawaii measurement. Claude's collapse holds across geographies, Gemma's does not. A category pattern from one geo turned out to be two once measured in two.
15-firm cohort. The Claude training-data collapse, found again in a second unrelated industry. Strong on the web-searching engines, near-zero on Claude. The competitive game is on the live-web tools.
12-firm AC-company cohort, 6,218 citations across 3 clean runs. The prediction was committed to a public timestamp before any data existed. Claude landed at 2%, the web-searching engines at 38-66%, Gemma at 63%. The third unrelated industry where Claude collapses on local firms, and the first one we called in advance. A forecast, not a case study.
The cross-category teardown reads all nine measurements against each other →
Every number we publish is on the record, with the run and the open-source code that prove it →
Which questions in your category get answered, who gets cited when you do not, and which kinds of sources the engines actually pull from when they decide.
Take one named reference. At Hawaii Theatre Center, a forensic readout surfaced what a standard scan walks past: a Charity Navigator profile not updated since 2023, a Better Business Bureau profile last touched in 1999, a missing Bing Business Profile, authority backlinks pointed at the wrong places. The quiet, citation-shaping detail nobody is looking at.
No bundled tiers, no per-seat math.
Free 1-page diagnostic before you commit. We run 5 real customer questions for your category across all 7 AI tools, then send a 1-page snapshot showing which competitors AI is naming and whether you’re one of them. One per business. The full engagement adds the rest of the question set (18 per category), weekly tracking, a cohort baseline that makes the numbers mean something, and a clear list of what to fix.
AI is becoming the front door, fast.
Sources: Google, OpenAI, Pew Research.
NeverRanked is $1,500 a month per market, $22,500 for the first year. A handful of new customers a year covers the whole cost. Every customer after that, named by AI because of the work, is money you would not have had otherwise. Pick your category.
One new case is worth, on average
It takes 2 new cases to cover a full year of NeverRanked. Every case after that is found money you would not have had.
Based on avg settlement of $40,000 to $55,000 with a 33% contingency fee.
We measure and diagnose. We do not promise customers, and AEO is an early space with no guaranteed result. This shows what one customer is worth to you, not a return we are promising. Being named by AI for the questions buyers ask is how you get found in the first place.
Does AI name your business when your buyers ask? Paste your URL and find out in seconds.
Run the free check →NeverRanked is a research practice, not a software company. The measurement, the memos, and the punch lists are produced by Lance Roylo, in Honolulu. There is no account layer between you and the person doing the research.
How the practice operates: we measure, we do not execute. We report what the AI engines actually cite, never what we claim our work caused. We do not promise a citation lift in advance. A finding that cannot be substantiated does not ship. That discipline is the product.
Lifted from the inbound emails Lance answers most often.
SEO measures search engine ranking factors on your own site. We measure what AI tools actually cite when buyers ask category-shaped questions, across 7 surfaces. The deliverable is also different: SEO tools give you a dashboard to interpret yourself; we hand off an interpreted research memo plus a prepped punch list your team executes. See /vs/ for the structural comparison.
Two reads on this. First, AI search usage in B2B and high-consideration consumer decisions is already non-trivial and growing on the curve we have public visibility into. Second, even if your buyer is not asking ChatGPT today, your competitor showing up there first when they do is the move you can not undo. We measure that surface so you know whether the move has already started.
We do not promise a lift in advance. The only promise we make is the measurement itself: you will know what AI cites for your category, what gaps exist, and what conditions a buyer of your category typically closes to move the needle. Whether your team executes the punch list well is what determines lift, and we measure that monthly so the answer is observable, not asserted.
We measure. We do not execute. We do not write content, edit pages, deploy schema, update profiles, or change your site. Your in-house team or your agency executes against the punch list we deliver. That separation is structural and is the whole position. It also means we never compete with your agency for execution hours.
Two ways. The instant self-serve check at check.neverranked.com tells you what the 7 AI tools can read from your site. The hand-built 1-page diagnostic runs 5 real customer questions for your category across all 7 AI tools and sends a snapshot showing which competitors AI names and whether you are one of them. One free diagnostic per business.
More: the full FAQ covers cancellation, NDAs, agency channel, data handling, and what happens if a finding turns out to be wrong.