BlogHow we work
Verify AI output before you act on it
Generated findings are a starting point. Here is the short list we check before a number, a competitor name, or a recommendation leaves the building.
AI can make mistakes. That sentence is on our reports because it is true, not because a lawyer asked for a footer.
Generated write-ups are fast at assembling a narrative from mixed evidence: crawls, listings, answer-engine excerpts, and whatever the model already believes about a category. Speed is the point. It is also how a wrong competitor name, a stale header, or a confident recommendation gets into a PDF.
We treat the model as a drafter. The human who sends the file owns the facts.
What we actually re-check
You do not need to re-audit the internet. You need to spot-check the claims that would change a decision.
Citation and mention rates. Recompute from the underlying results if the summary feels too round. A “strong AI presence” that is only branded prompts is a different claim than presence on category questions. If the Index and the rate disagree, look at featured vs mentioned before you explain it to a client.
Competitor names. Models merge similar businesses, revive closed listings, and invent plausible LLCs. Search the exact string. If you cannot find the firm in the same city, do not put it in a slide.
Sampled URLs. Open the pages the crawl says it used. Confirm the title, the H1, and that you are not looking at a staging host or a parked domain. Canonical tags that point off-domain are a frequent source of “the site said X” when the site did not.
Security and header findings. If a report says a header is missing, fetch the live response. CDNs and preview hosts disagree with what a scanner saw an hour ago. Do not send a “fix HSTS” ticket until you have seen the headers yourself.
Recommendations. Treat them as a backlog candidate list. “Add FAQ schema” is only useful if the page is a candidate for FAQ and you want that rich result. “Write more content” is not a finding.
What we do not pretend to know
We do not know next month’s AI Overview mix. We do not know whether a mention caused a phone call. We do not know that a Citation Index of 72 “means” anything in isolation — it is a weighted read of the prompts you ran, on the engines you queried, on that day.
If a client asks for a guarantee of rankings or citations, the honest answer is no. Research output is not a contract with Google or OpenAI.
A working habit
Build a five-minute pass into delivery:
- Read the summary out loud. Any sentence you cannot point to in the data gets cut or rewritten.
- Click three sampled URLs and one “competitor” name.
- If AI Overviews were part of the scan, look at those rows separately. Zero citations there is a specific finding, not a rounding error.
- Keep the disclaimer on the file. It is for the reader, including you, two weeks later.
The terms say the same thing in legal language: you are responsible for verifying important findings before you act on them. The operational version is shorter. If you would not defend the sentence on a call, it does not ship.