Check what AI says about your business by asking ChatGPT, Gemini, and Google AI answers the same questions buyers ask. Record each prompt, answer, source link, correct fact, error, and omission. Then trace every problem to the page responsible for that fact and repair the visible copy, metadata, structured data, discovery files, or internal links at the source.
You can run a useful first audit today with a written question list and a simple worksheet. No tooling, no subscription. The goal is to see what buyers are being told, find the source of each problem, and walk away with a specific repair list.
Answers change across engines and over time, so date every observation and keep the prompt wording fixed while you compare.
Write the buyer questions first
List the questions a buyer asks before choosing your business. Use ordinary language and put the business or offer name inside each prompt.
Start with questions such as:
- What does this business do?
- Who is this service for?
- Does this business provide the service I need?
- What does the service include?
- How does the working process operate?
- What does the buyer need to provide?
- Which limits or requirements should a buyer know?
- How can a buyer start in writing?
Then mine the real material: written inquiries, sales notes, support messages, existing page headings. Give each question one primary page that owns the answer.
Run the same test across the engines
Open a fresh session in each engine so earlier prompts cannot leak context into the answer. Ask the question exactly as written. No leading language. If you teach the engine the answer you want, you are auditing your own prompt.
Run the audit in ChatGPT, Gemini, and whichever of Google's AI answer surfaces you can reach. Record the engine and the surface, because different interfaces return different source links and formats.
If an answer asks for more context, write that down too. It usually means the business name, offer, or question is ambiguous on the open web, and that is a finding.
Record the complete observation
One row per question per engine. Capture the answer word for word so later reviews do not run on memory.
| Field | What to record |
|---|---|
| Question | The exact buyer prompt |
| Engine and surface | Where the question was asked |
| Review date | When the answer appeared |
| Answer | The complete response |
| Source links | Every cited or linked page |
| Correct facts | Statements supported by the business |
| Wrong facts | Statements that conflict with supported information |
| Missing facts | Information needed for a complete answer |
| Primary owned page | The page responsible for the answer |
| Repair | The specific page or machine-readable change to make |
Screenshot anything where the interface makes the source relationship hard to preserve. Store it with the worksheet under a filename with the question label and date.
Classify what came back
Give each observation a plain status:
- Accurate and linked to the responsible owned page
- Accurate with an incomplete source trail
- Partly accurate with missing context
- Stale or contradictory
- Unsupported by any source you can verify
- No useful answer returned
Then mark each incorrect sentence individually. A broad "bad answer" label leaves you nothing to fix. A sentence-level note points at the outdated service description, the missing boundary, or the conflicting third-party page.
Trace errors to the source
Open every source the engine cites. Check whether the wrong fact appears there, whether the page has changed since, and whether the answer dropped an important condition.
Then inspect the page you control:
- Does the page answer the question directly in visible HTML?
- Does the answer name the business, offer, and relevant condition?
- Do the title and description describe the same page purpose?
- Does structured data match the visible fact?
- Does the canonical URL point to the approved page?
- Do related pages link to it with descriptive text?
- Does the sitemap include it?
- Do discovery and answer files point to the current version?
A missing answer usually traces to a missing passage or page. A wrong answer traces to stale copy, conflicting owned pages, or a third-party source. A vague answer usually means the service description is vague, and the engine is just being honest about it. An absent page points at a crawl, indexing, or internal-link problem.
Repair one fact at its primary page
Write the approved fact in a direct answer near the top of the responsible page. Add the detail and evidence a buyer needs. Then bring the metadata, structured data, canonical reference, internal links, sitemap, discovery files, and answer-file summaries into line with it.
Update every other owned page that repeats the fact. Keep one approved wording in the fact map and note every surface that uses it.
For a wrong third-party page, record the source and the correction needed. Use whatever correction process you have authority over for that listing or profile. Either way, strengthen the owned page so the accurate version stays clear and current.
Repeat the audit without changing the test
Same prompts, same engines, same surfaces, same worksheet fields after the repair. New review date, earlier observation preserved. Now you have a comparison the business can actually inspect instead of a feeling that things got better.
Run it again when the offer changes, priority pages move, public profiles change, or a buyer reports a surprising answer. New buyer questions go on the standing list as they show up.
Improving the owned source gives engines clearer material to use. It cannot guarantee a ranking, citation, recommendation, or answer. Nobody honest sells those.
Keyframe0 offers a free written site review. Send the web address through Work with us to get what stands, what is missing, and what to fix first.