sample audit
Sample audit: before and after, on somebody’s real website
A trade contractor, five facts, one install. The baseline was frozen before the block existed; the same script, the same five models and the same day were used for both runs. The client’s name is withheld until they choose to be named.
Method, in five lines
- Five models were each sent to the site cold, with a fetch tool and no hints, and asked buyer questions.
- An answer counted only if the model stated the fact and attributed it to this business. Near-misses counted as misses.
- The baseline ran and was frozen before anything was installed.
- The block carried five facts, all of them already published somewhere on that site. No price, no lead time, no warranty — the client has never published those, and we do not invent commitments.
- The after-run used the identical script on the identical questions.
What the agents recovered
| Buyer question | Before | After | |
|---|---|---|---|
| How long have you been in business? | 1/5 | 5/5 | shipped in the block |
| Do you do free estimates? | 0/5 | 4/5 | shipped in the block |
| Are you licensed? (licence number) | 5/5 | 5/5 | already at ceiling |
| Where do you work? | 5/5 | 5/5 | already at ceiling |
| What is your phone number? | 4/5 | 4/5 | already at ceiling |
| What does it cost? | — | — | deliberately not published; still reported unavailable by all five, before and after |
| How long is the wait? | — | — | deliberately not published; unchanged |
Measured 2026-07-29 on a live site, five models per arm, one run. Source: MARKETING_NOTES_AGENT_VISIBILITY.md, Note 30.
Both facts that were genuinely missing moved sharply. Both gaps we deliberately left empty stayed exactly shut. That is the most useful thing on this page: a partial fix produced a partial result with no halo, which is how you can tell the measurement is not flattering the intervention.
The number a business owner should care about
| Measure | Before | After |
|---|---|---|
| Mean pages fetched per visit | 2.4 | 1.0 |
| Deepest model, pages fetched | 8 | 1 |
| Mean stated confidence | 0.86 | 0.85 |
Same run, Note 30. The 2.4→1.0 average is not a broad drift: it is one model going from eight fetches to one.
Before the install, exactly one model in five worked hard enough to assemble the picture, crawling five pages to do it; the other four read the homepage once and answered incompletely. After, every model got the facts on the first fetch. You are not buying more attention from agents. You are buying the right answer inside the one page they were always going to read.
Limits of this audit — read them before you quote it
- Five models per arm, one run, one site. Not a population estimate.
- Three of the five shipped facts were already at ceiling, so the movement rests on two of them.
- Stated confidence did not rise (0.86 to 0.85). The block does not make agents more certain, only more correct.
- The depth collapse belongs mostly to one model. It reproduced a similar effect seen in a controlled rig (Note 19), and it did not reproduce on a later client site (Note 32) — a site-dependent effect, not a law.
- We formed a hypothesis that a partial block would suppress the recovery of facts it omitted, tested it against the actual fetch paths within ten minutes, and killed it. Nothing got worse. It is recorded because the next partial install will raise it again.
Your audit
Yours will not look like this one. It will have your questions, your ceiling facts and your gaps, and the “before” will be frozen before anything is installed. If the facts you are missing turn out to be already recovered at ceiling, we will tell you that, and you should not buy the package.