How we measure AI retrieval of pricing records
The AI Retrieval Index counts how often AI assistants retrieve CostBench’s software pricing records. This page defines every number on that surface, how it is collected, over what window, and — most importantly — the limits of what it can tell you. Data currently refreshed July 12, 2026.
Three separate signals — never blended
We track three distinct numbers. They are always labelled and reported separately. We never add them together, and we never call any of them a “citation.”
user-initiated retrievals
The headline signal.A live fetch of a product’s CostBench pricing record made by an AI assistant while it answers a real user question. These are the requests that matter for “how often do assistants research this pricing.”
Detected by user-agent: ChatGPT-User, Perplexity-User, and OAI-SearchBot (the on-demand fetchers assistants use when a user asks something).
crawler fetches
Not user demand.Background index crawls — the bots that build and refresh AI training and search indexes on their own schedule, with no user in the loop. Reported separately and never counted as user activity.
Detected by user-agent: GPTBot, ClaudeBot, Amazonbot, Bytespider, PerplexityBot, and similar background crawlers.
AI-referred visitors
Humans, not machines.Distinct people who arrived on a pricing record after clicking through from an AI assistant. This is human traffic that an assistant sent onward — a different thing from the machine retrieving the record.
Detected from referral analytics: an ai_referral event fires when a visit’s referrer is an assistant (chatgpt, perplexity, claude, and others).
How the data is collected
Every request to CostBench carries a user-agent string. We classify each request server-side into one of the buckets above by matching that user-agent against the known assistant fetchers and background crawlers. User-initiated retrievals and crawler fetches both come from these server logs; AI-referred visitors come from referral analytics on human page views. A request is only counted once, against exactly one bucket.
Every figure is scoped to a single product’s CostBench pricing record. It measures how often assistants retrieve our record for that product — a concrete, honest proxy for research demand — not the product’s standing anywhere else on the internet.
Window and refresh cadence
Eligibility. A product is marked eligible when it clears
user_initiated_retrievals_30d >= 25 OR ai_referred_visitors_30d >= 5. Below that threshold there simply isn’t enough AI activity to
say anything useful, so we say so rather than dress up a small number.
What a retrieval count does not mean
A retrieval is a specific, narrow event. It is easy to over-read, so here is what it is not.
- A record fetch is not an answer citation. When an assistant retrieves our record, we cannot see whether it used our data in its answer, credited us, or quietly discarded it. We count the fetch, which we can observe — not the citation, which we cannot.
- Neither is it brand visibility. A high retrieval count says assistants often reach for our pricing record for that product. It does not measure how visible the product is across the whole AI ecosystem, how often it’s mentioned, or how it’s ranked.
- Retrievals are not buyers, and not clicks. Machine retrieval and human referral are genuinely different — many heavily-retrieved records send almost no humans onward. That’s why AI-referred visitors are reported as their own number and never folded into the retrieval total.
- Crawler fetches are not user demand. Background crawlers run on their own schedule regardless of whether anyone is asking about a product. We keep them out of the user-initiated number entirely.
- This is a single window, not a trend. We publish a point-in-time snapshot. We do not make “rising” or “falling” claims until we have a comparable prior window to measure against.
In short: this is an honest measurement of how often AI assistants retrieve CostBench’s pricing record for a product — a useful proxy for research demand, and nothing more than that.