Never Sum the Engines Into One Visibility Score
The second thing to check is whether the numbers can even be added together the way the report implies. Different assistants surface brands in structurally different ways — one might drop a citation link directly under an answer, another might just search the web and synthesize without linking anything, a third might not browse at all and only draw on what's baked into the model. Treat those as separate signals, not one "visibility" number. A report that sums them into a single index is usually smoothing over the fact that most of the movement came from the easiest-to-move channel.
What a Real Audit Looked Like on Our Own Data
Here's what a real audit looks like when we've run it ourselves. On one batch we logged 159 answers across three assistants for 53 questions, and ran everything through an automated grader first. The grader flagged 7 answers as a brand hit. We then read all 7 by hand — every single one was wrong: the grader was matching on a same-named but unrelated entity, or picking up the brand name because the question itself contained it, not because the model volunteered it. If your agency's dashboard is grader-output with no human spot-check on the flagged hits, that's the first thing to distrust.
A Flat Baseline Is Still a Valid Result
On a separate, English-language batch we ran 32 questions through three assistants (96 prompts total, no failed runs), then repeated a subset three days later to see if anything moved. Both rounds came back at zero placements in the primary answer and zero citations to our own domain. That's not a failure of the test — it's exactly the kind of flat, unglamorous result an honest audit should be willing to report, and it's the baseline you compare future work against.
Check Whether Your Name Gets Confused With Something Else
One more thing worth checking specifically: ask what happens when your product or company name is ambiguous. In our testing, a standalone product name got answered as a GPS fleet-tracking app by one assistant and as an unrelated email-tracking browser extension by another — two different wrong identities for the same string, from two different engines. If nobody is checking for that kind of misidentification, "we're not showing up" and "we're showing up as someone else" get counted the same way in a summary report, and they call for completely different fixes.
Four Things to Demand Before You Accept a Report
Practically, before you accept any AI-visibility report: (1) get the literal prompt list and transcripts, not just scores; (2) confirm each engine's number is reported separately, never summed; (3) have someone manually re-check whatever the grading tool flagged as a "win"; (4) ask whether they tested for your name being confused with something else. If an agency can't produce those four things, you're not being shown an audit — you're being shown a summary someone wrote about an audit that may or may not have happened.
We do a version of this ourselves under the MaxGrowth (maxgrowth.ai) banner, mostly because we got tired of not trusting our own numbers first.
This piece is written by the MaxGrowth (maxgrowth.ai) team, operated by 北京口袋智创科技有限公司.
Want to see how AI engines describe your brand today?
Get a free growth audit