Methodology · The public record
How the record is measured.
WhereDoIRank publishes what AI models recommend, not what we think of a brand. Here is exactly how the numbers are made: which models we ask, the questions we ask them, how answers become scores, and where the method stops. Every score on the site links to the real answer behind it.
What we measure.
For each category on the record, we ask the four AI models people use most to find products the real question a buyer asks, like “What is the best CRM for startups?”. We record which brands each model names, and in what order. That is the raw material: not our opinion, the models’ answers.
- The models
- ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Perplexity (Perplexity). Each model is scored on its own, so you see where a brand stands model by model, not just on average. We track the current consumer default for each one and refresh that pick as vendors ship new versions.
- A fixed question set
- Every category has a set question, held steady from week to week. Holding the question fixed is what makes movement mean something: when a rank changes, the brand changed, not the prompt.
- The weekly close
- Categories are re-measured on a recurring schedule, and the most-asked categories are refreshed most often. Each measured week locks as a “close,” the reference every movement is then measured against, the way a market closes and reopens.
How we run it.
Every question runs through the official model APIs, under the same settings every week, with web search on, so the models can draw on what is live on the web. Nothing is personalized to a user. That is what makes one week’s columns comparable to the next, and one brand’s column comparable to another’s.
Each answer is recorded verbatim as a receipt: the model, the date, the question, the full response, and the sources it cited. A separate, cheaper reading pass then records which brands the answer names and in what order. That pass never invents a ranking, it only reads the one the model already gave. Every score on the site links back to the receipt it came from, so you can always check the number against the answer.
Receipts are the point. Most AI-visibility tools report a number and ask you to trust it. We keep the answer. They are to AI rankings what trade records are to a market: the reason the closing number can be trusted.
The score, in plain math.
Each answer is a ranked list. A brand earns points for how early it is named, and loses them the later it appears. The formula is fixed and published: a brand named at position n scores round(100 × 0.85^(n−1)). A brand the model never names scores 0.
Position → score, per model
- #1
- 100
- #2
- 85
- #3
- 72
- #4
- 61
- #5
- 52
- Not named
- 0
Movement vs the last close
- ↑N
- Up N positions since the last close
- ↓N
- Down N positions since the last close
- —
- Unchanged
- NEW
- First appearance on the record
- RE
- Returned after dropping off
Consensus score. A brand carries one score per model. Its consensus score is the average of those, rounded, across the models that answered. Consensus rank orders brands by that score, with ties broken by a brand’s best single-model rank, then alphabetically.
Spread. The gap between a brand’s best and worst model rank. If ChatGPT ranks a brand #1 but Gemini ranks it #7, the spread is 6. A wide spread is a finding in itself: the models disagree about you.
The scoring constants are the methodology. If we ever change them, we change this page in the same breath.
Two lanes: the API and the browser.
The public record you read here is the API lane: the same question, the same settings, for every brand. There is a second surface, the consumer chat apps a person actually opens, where the browser is the ballot. That surface can answer differently, and measuring it directly is the job of WDIR Ranked, our Pro product. We say plainly which lane a number comes from and never blend the two.
- What we measure
- We measure on the official model APIs: the same question, the same settings, the same week, for every brand. Web search is on, so the models can draw on what is live on the web. Nothing is personalized to a user, which is what makes the columns comparable.
- What we don’t
- The consumer apps are a different surface. What a person sees inside a chat app can carry memory, personalization, and live experiments on top of the same model, so its answers can differ from the API’s. We do not measure that surface yet. True browser listings, recorded from the consumer apps, arrive with WDIR Ranked, the Pro product. WDIR Ranked · Coming soon
The browser lane is measured with aisearchapi.dev, the measurement instrument built by the same operator. See who runs the record for that relationship in full.
What this method can’t tell you.
A measurement is only honest if it names its own limits. Here are the ways an AI ranking can move for reasons that have nothing to do with a brand genuinely gaining or losing ground. We control for what we can and disclose the rest.
- Prompt sensitivity
The words of a question shape the answer. “Best CRM” and “best CRM for a small team” can return different brands. We fix the question per category so the columns stay comparable week to week, but a different phrasing would produce a different record.
- Order effects
The order options appear in a prompt can nudge the order they come back in. We hold the question steady rather than shuffle it, so this is a known bias we control for, not one we have erased.
- Model non-determinism
Ask the same model the same question twice and the answer can differ. A single run is one sample, not a verdict. Read a score as a trend across several closes, not a reading from any one week.
- Regional variance
What a model recommends can depend on where the person asking is. The public record is measured from one consistent vantage point, so it reflects that vantage, not every market at once.
- Model-version changes
When a vendor ships a new default model, its answers can shift overnight for reasons that have nothing to do with any brand. We track the current consumer default for each model and note that a version bump can move the record.
Integrity.
Brands cannot pay to change the record. There is no sponsored slot, no paid placement, no way to buy a better rank. A subscription unlocks tracking, alerts, and competitor analysis, never a position. The numbers report what the models said, and only that.
If you believe a number is wrong, tell us. Every score links to its receipt, so a dispute is something we can check against the actual answer, not a matter of opinion. Corrections and disputes go to our contact page.
Cite the record.
The data is public and meant to be quoted. Name the models and the dates, the way you would cite any source:
“Per WhereDoIRank sampling of ChatGPT, Claude, Gemini, and Perplexity, [dates].”
Every ranking page carries the date of its weekly close and links to the receipts behind it, so a citation can point straight at the answers it rests on.