A growing share of buying research now happens inside an AI assistant, and unlike a search result you cannot see your position. The only way to know what is being said about you is to ask.
This is a repeatable test you can run in an hour and repeat quarterly. The findings are usually uncomfortable and always useful.
- 1
Ask directly, in several assistants
Ask ChatGPT, Claude, Gemini and Perplexity "What is [your business name] in [your town]?" Use a fresh chat with no prior context, since an assistant that has seen your website earlier in the conversation will simply repeat it back. Record each answer verbatim.
- 2
Ask the questions a customer would ask
More telling than asking about you by name: "Who are the best [your service] companies in [your town]?" and "Who should I use for [specific problem]?" If you are not mentioned, that is the real finding. Being described accurately when named is a much lower bar than being recommended unprompted.
- 3
Note every factual error
Wrong services, wrong location, wrong size, confusion with a similarly named business, out-of-date information. Write them down verbatim. Errors repeated across several assistants usually trace back to one incorrect source they have all drawn from.
- 4
Find out where the answer came from
Perplexity cites sources directly, and ChatGPT and Gemini often will if asked. Follow those links. It is common to find the answer built from a directory listing you have never updated rather than from your own website.
- 5
Fix the sources, not just the site
If a stale directory entry is driving a wrong answer, correcting your own website will not fix it. Update the source. This is why consistent listings across directories, trade bodies and social profiles matter more for AI answers than for traditional search.
- 6
Fill the gaps the answers reveal
If an assistant cannot say whether you cover a town, or what you charge, or whether you handle a particular job, that information is missing or unclear on your site. Each gap is a page or a section worth writing — and you now know it is being asked for.
- 7
Re-run it quarterly and keep the answers
Keep a document of each round with the date. Models update constantly, and the trend over a year is far more informative than any single answer. It also demonstrates whether the work is landing, which is otherwise very hard to show.
Why answers differ so much between assistants
Some search the live web at the moment you ask; others answer from training data that may be many months old; some do both. A business can be described well by one and be invisible to another. Test several, and do not over-react to any single result.
What actually moves the needle
Being mentioned across multiple independent sources, having clear factual content on your own site, and consistency between the two. There is no way to submit a correction to a model — you change what it reads.
Frequently asked questions
There is no general correction mechanism. The practical route is to fix the underlying sources — your site, your listings, anywhere inaccurate information is published — and wait for the models to pick up the change.
Would rather we did it?
These guides are here so you can do it yourself. If you would rather hand it over, that is what we do.
Start a conversation