Check what AI search says about your business.
A brand mention is easy to screenshot and surprisingly easy to overread. Here’s how I’d check an answer before deciding what needs fixing.
Ask something a customer would actually ask.
Start with a need that brings people to your business before they know its name.
Sales calls, enquiries and support questions are useful here. Choose a customer type and a problem. Add the constraints that affect the choice: a city for a local service, an existing tool for software, or a budget if it matters.
I’d start with a few questions you recognise from the work. You can finish a small first pass in about half an hour. It won’t represent every possible customer, but it will give you something specific to investigate.
- Discovery“Which bookkeeping services suit a small design studio using Xero?”
- Comparison“What should a design studio compare when choosing a bookkeeper?”
- Branded fact check“Does [business name] support Xero for design studios?”
Keep the branded check separate. Once you put your name in the question, you’re testing what the system says about you when asked.
Make the next check comparable.
Save the wording and the conditions. You’ll need them when another answer disagrees.
Choose a search surface to begin with. Record the engine and whatever model or mode label the interface shows. If it doesn’t show one, write “not shown.” Guessing adds detail without adding information.
For a web search check, enable search where you can. Note whether it was enabled, visibly used, disabled or unclear. A response without citations doesn’t establish that search never happened.
Use a fresh conversation for each question. Note the language, market, sign-in status and any memory or personalisation settings you know about. A fresh chat reduces earlier conversation context; account settings can still matter.
Save the answer excerpt and its source URLs. In ChatGPT, sources may appear as inline citations or in a Sources view. Open them. OpenAI’s search guidance also explains how location and memory can affect a search.
If a Google AI Overview doesn’t appear, record “AI answer not shown” and leave the answer signals unassessed. There wasn’t an answer in which to look for your brand.
Check what the answer actually supports.
A name, a recommendation and a source link tell you different things. The worksheet keeps them separate.
| Signal | What you can record | What still needs checking |
|---|---|---|
| Brand mentioned | The business is named, with or without a link. | Whether the description is accurate and the mention is favourable. |
| Recommended | The answer suggests the business as an option. Keep qualifications such as “worth comparing.” | Whether the suggestion fits the customer’s requirements. It tells you nothing about their eventual choice. |
| Source cited | The interface attaches a source URL. Note whether it’s your site or a third party. | Whether the page supports the claim and is current. A citation alone doesn’t establish a visit. |
| Not observed | The signal didn’t appear in this particular answer. | Whether it appears for another question or under different conditions. |
Use “Not assessed” when you haven’t checked. An empty field shouldn’t quietly become a “No.”
Read around the name
“Avoid this provider” still contains a brand mention. “Could be worth comparing” is a qualified suggestion. Keep the wording so someone else can see why you recorded it that way.
Open the citation
Check who publishes the page and which claim it supports. A domain written in plain text is an unlinked reference; record it as such rather than counting it as a clickable citation.
Choose something you can investigate.
An odd answer gives you a place to look. It rarely tells you why the system got there.
Verify the fact on your own site
If the service, location, price or eligibility is wrong, find the current primary source and save it alongside the answer. Check what your page actually says before assuming another source caused the error.
Inspect the page being cited
An old directory entry or incomplete service page is worth checking. Correct information you control and request corrections where you don’t. Make the page useful to the customer who has that question.
Repeat the original question
Keep the setup comparable. If you change the wording, market or mode, save a new question variant. A different answer after an edit is worth recording; a small uncontrolled audit won’t establish that the edit caused it.
Start with the actual information
Explain the service, customer fit and relevant constraints in accessible page text. Google’s guidance for AI features says its usual SEO practices apply and special AI markup isn’t required. Meeting the requirements doesn’t guarantee inclusion.
Check whether an automated reader can get that text in the first place. I built the AI Readiness Audit to inspect crawler rules, server responses and the content available before and after JavaScript runs. Test the actual page you care about. Passing an access check doesn’t establish that a search system has crawled or will cite it.
When manual checks become a chore
Keep a stable question set and dated answers. If that gets difficult to maintain, a tracking tool can help. I contribute to Beamtrace, the AI visibility tracker we’re building at Elfsight. It’s a commercial option for organising prompts and monitoring brand presence; check the current engine coverage against what you need.
The worksheet here works on its own. My involvement in Beamtrace is worth knowing when I mention it. The product pages below explain its different monitoring views.
Give the tracker a question to answer.
Once you’re keeping a regular record, choose the view that helps with the work in front of you. These are product pages for Beamtrace, the commercial tracker I contribute to at Elfsight. Check the current coverage and how each metric is calculated before relying on it.
Search visibility tracker
For checking whether the business appears across your question set. Keep the questions, engines and dates behind a summary score available to inspect.
Search rank tracker
For following placement within tracked answers. Read the wording around the position: an ordered shortlist and an unranked paragraph need different treatment.
Competitor analysis
For seeing which other businesses appear for the same questions. Compare the same customer need and conditions, then inspect the answers that account for the difference.
Brand mention tracker
For following when the name appears and what is said about it. Keep the excerpt; a favourable recommendation and a warning can both contain the name.
Citation analysis
For inspecting linked sources and patterns across answers. Open the actual page and check the claim it supports, including third-party citations.
AI reputation management
For looking at how the business is described alongside customer-review themes. Verify specific complaints or claims against their sources before choosing a response.
Before you rewrite the page
I publish research briefings that fit the checks above. Each briefing links to its source paper and discusses what the study can support.
Check the language you’re testing in
A small ChatGPT study on accounting software shows why the question’s language belongs in the record. Its result is specific to that setup; I’d use it to plan a check in the customer’s language.
Read the source paperQuestion the promised rewrite
The SAGEO Arena benchmark tests retrieval, reranking and answer generation together. A rewrite intended to improve the final answer can hurt an earlier stage. Read the setup before applying a reported gain to your own pages.
Read the source paperKeep visits separate from mentions
A Google Search field experiment looks at publisher click-through when AI features change. It gives a reason to measure referrals alongside answer visibility. A mention count won’t tell you whether anyone visited.
Read the source paper
After someone reaches your website
The search answer is only part of the visit. People still need to check what customers say and get an answer to the question that brought them here.
Show customer reviews with JustReviews
Bring existing reviews onto your site. Read them with their dates and original sources; a customer’s experience needs more context than an average star rating.
Answer visitor questions with a website chatbot
Use your website content, files and FAQs to answer customer questions. Check the answers against current prices and policies before leaving the chatbot to handle them.
Sources & notes
Reviewed 5 October 2026.
Products and interfaces change; the source pages are linked here so you can check the current details.
- Google Search Central: AI features and your website
Eligibility, technical requirements and the limits of inclusion.
- OpenAI: Searching the web with ChatGPT
Search, sources, location and memory.
- The worked example
Northline Books, Studio Ledger and their answers are fictional.
This is my independent publication, part of a website publishing experiment. I contribute to Beamtrace at Elfsight. The audit tool and research briefings linked here are also my work. Those connections are worth knowing when you read my recommendations. The manual audit is an editorial recommendation, not a provider’s official visibility score. AI assisted with drafting and implementation.