We simulate the conversations buyers already have with AI. Designed buying scenarios run repeatedly against Claude, GPT, and Gemini, and two separate analysts score the results without seeing each other's work. The audit finds the leverage points that win recommendations, then keeps measuring whether they moved.
The controls
Four safeguards inside every audit
The home page shows the steps. These are the safeguards each step runs under, so no finding can be bent to fit a conclusion.
01 / RAW CAPTURE
Saved before anyone reads it
The model's answers are written to disk as they come in. Nobody summarizes or cleans them first.
02 / BLIND SCORING
Two scorers, no contact
One scorer grades how the model talks about you. A second checks whether it got the facts right. They never see each other's notes.
03 / PROVENANCE
Every number has a source
Click any figure in the report and you land on the exact model response it came from.
04 / CONTROL GROUP
Some scenarios we never touch
A few buying scenarios never get any new evidence. If those move too, the market moved, not you.
Common findings
The findings we see most often
AI is repeating what others said about you
With no first-party evidence to read, the model answers from training memory and third-party reviews. Nobody at the brand has seen or corrected that material.
What it sounds like"Reviewers tend to describe them as a lighter-weight option; for enterprise I would look at [competitor] first."
The model will tell you what would change its mind
Every conversation ends by asking what evidence would flip the recommendation. Collected across runs, those answers are the work plan.
In the model's words"If independent security verification (e.g. SOC 2) were publicly documented, I would include them on this shortlist."
Your site blocks the crawlers that would cite you
One AI company we audited had published an llms.txt for the same bots its robots.txt turns away.
What we foundrobots.txt → ClaudeBot: blocked · GPTBot: blocked llms.txt: published, unreachable to both
A mention with a caveat is not a recommendation
"Depends on your size" kills buyer confidence, and it still counts as a mention on every visibility dashboard.
A hedge that counts as a mention"[Brand] could work here, though it may depend on your team's size and how mature your data practice is."
Our Perspective
Being mentioned is not being chosen
AEO replays the SEO playbook: buy prompt data, optimize to rank in answers, track a weekly score. That tells you whether AI mentions you. A buyer needs AI to choose you, which is a different thing to measure. Here is where the two part.
Rewording moves the score
The dashboard measures phrasing, not performance. We swapped one synonym across a hundred buyer prompts, 600 runs in all. Share of voice moved 6.1% on average and 16.7% at worst. That is as much as a real brand investment moves it.
Best tools for revenue forecasting?
Anaplan, Pigment, Clari
Best software for revenue forecasting?
Anaplan, Pigment, Workday
one word changed and Clari dropped out
Nobody can check the sample
The vendor picks the prompts and the competitors that set your score. The prompt data comes from opt-in browser panels, and no major platform publishes an accuracy figure for its headline metric.
34%
share of answer
prompt setchosen by vendor
competitor setchosen by vendor
sampleopt-in browser panel
accuracynot published
Models discount the content
Content written to be cited reads like the material models distrust. In our audits they say so unprompted.
Is Northwind SOC 2 certified?
Their site says so, but those figures are self-reported, not independently confirmed.
said unprompted in an audit conversation
GEO asks whether AI mentions you. We ask whether AI chooses you.
GEO / AEO
AI Brand Alignment
Object of work
Citations and visibility
The model's beliefs and verdicts
Unit of analysis
Pages, snippets, mentions
Full buying conversations
Stage covered
The search-shaped moment
Framing through final call
What it measures
Presence
Conversion and objections
Output
A visibility dashboard
A prioritized evidence brief
Fails when
You are visible but not chosen
It measures the choosing itself
Find out what the models believe about you
A baseline audit names what AI gets wrong, what it hedges, and the evidence that would change its mind.