On this page
Most founders find out they have an AEO problem by accident. You ask ChatGPT for the best tool in your category, your product is not in the list, and four competitors are. An LLM visibility checker turns that accident into something you can measure and work on. This is the checklist I use to set one up, read the output, and decide what to change first.
What an LLM visibility checker actually gives you
A checker runs a set of buyer questions through assistants such as ChatGPT, Claude, Perplexity, and Gemini, then reads the answers looking for your brand. The output is a rate: in how many of those answers you were named.
That single number is less useful than the breakdown behind it. A checker worth the setup time separates:
- How often you appear, per assistant, not averaged into one figure
- Where in the answer you appear, since being named fourth in a list is nearly invisible
- How you are framed, because “limited compared to X” and “good fit for small teams” are both mentions
- Which URLs the assistant pulled from when it answered
The last one is the part founders underuse. Mention rate tells you the score. The cited sources tell you which pages you need to exist on, which is the closest thing AEO has to a to-do list. That is why citation tracking is worth setting up even before you have a visibility trend to look at.
1. Write the prompt list your buyers would actually type
Skip your own vocabulary. Use the phrasing a buyer uses before they know your product exists. Four kinds of prompt cover most of the ground:
- Category prompts. “Best time tracking tool for agencies”, “tool to monitor uptime for a small SaaS”.
- Comparison prompts. “Alternatives to [competitor]”, “[competitor A] vs [competitor B]”.
- Problem prompts. The symptom, not the product: “how do I stop double-booking clients”.
- Branded prompts. “What is [your product]”. Only two or three of these, as a fact check.
Branded prompts rarely move your score, but they catch a separate failure. If an assistant confidently describes pricing you retired a year ago, that is a correction job, not a visibility job.
2. Run every prompt more than once
Ask the same assistant the same question three times and you can get three different lists. Responses are sampled, not retrieved from a table, so a single pass is a coin flip reported as a fact.
This is where most one-shot graders break down. Repetition fixes it: run each prompt several times and report a percentage. “Named in 8 of 12 runs” is a measurement. “Named” is an anecdote.
Two practical rules:
- Sample each prompt at least three times per assistant. Past five, you are paying for precision you will not act on.
- Keep the prompt text frozen between runs. Rewording a prompt and comparing to last month measures your editing, not your visibility.
On volume, 20 to 30 prompts is a reasonable floor before week-to-week movement means anything.
3. Check more than one assistant, and expect them to disagree
Providers fail differently, which is the single most useful thing a multi-model checker reveals.
| Assistant type | What it rewards | Typical failure |
|---|---|---|
| Retrieval-heavy (Perplexity, AI Overviews) | Fresh, crawlable, citable pages | You are invisible to the crawler |
| Training-heavy (ChatGPT, Claude without browsing) | Being an established named entity | Nobody outside your site talks about you |
Seeing 45% on Perplexity and 4% on ChatGPT is common, and it is not a contradiction. It means your pages are reachable but your brand is not established in the wider corpus the model learned from. The reverse split points the other way: people mention you, but a crawler cannot read your site.
An averaged score hides this and sends you off fixing the wrong half.
4. Confirm you are reachable before you write anything
Before adding content, rule out the boring failures. A brand can be entirely absent from AI answers for mechanical reasons:
- AI crawlers blocked in robots.txt, sometimes by a rule someone added years ago. Start with allowing AI crawlers explicitly.
- Core pages that only render with JavaScript, so a fetcher sees an empty shell.
- No plain definition of what your product is and who it is for, in text, on a page a model can quote.
- No presence anywhere but your own domain: no listings, no comparison pages, no third-party mentions.
A product with none of these covered has nothing to retrieve. Publishing more blog posts in that state changes very little.
5. Put your competitors on the same prompt list
Your own rate in isolation does not tell you whether 18% is bad. Running the identical prompts against the products that keep appearing instead of you does.
What you are looking for is not the gap in the headline number. It is the shape of it:
- Prompts where a competitor appears and you never do. That is a content gap with a specific topic attached.
- Prompts where you both appear but they are named first. That is a positioning and framing problem, not a coverage problem.
- The sources cited when they win. Often the same handful of roundups and directories, which becomes a concrete outreach list.
The mechanics of setting this up are covered in the guide on tracking competitors in AI answers. The habit matters more than the tooling: check it monthly, not daily, because visibility moves slowly and daily numbers are mostly noise.
6. Make it a loop, not an audit
A one-time check tells you where you stand. It cannot tell you whether anything you did helped. That requires the same prompts, the same sampling, and the same assistants, repeated on a schedule.
A monthly rhythm that holds up:
- Re-run the frozen prompt set. Do not edit it this week.
- Note the per-assistant rates and the direction of each.
- Pick the two lowest-hanging gaps from step 5, not ten.
- Ship fixes: a definition page, a comparison page, a listing, an outreach email to a roundup that already ranks.
- Wait a full cycle before judging the result.
The waiting is the part people skip. Retrieval-based assistants can pick up a new page within days. Assistants leaning on training data may not reflect your work for months, and no amount of refreshing the dashboard changes that.
Start with one afternoon
You do not need a tool to get a baseline. Pick ten prompts using the categories above, run each three times in ChatGPT and Perplexity, and write the results into a spreadsheet: prompt, assistant, mentioned yes or no, position, who else was named. That is sixty runs and roughly an afternoon, and it will tell you which half of the problem you have.
Automate it once you want the trend instead of the snapshot, since repeating that spreadsheet by hand every month is where the habit dies. AskAiRank runs the same prompt set on a schedule across the major assistants and keeps the competitor comparison next to your own numbers, which is the version of this checklist that survives a busy month.
Frequently asked questions
Answers about running an LLM visibility checker and acting on what it reports.
Plan on four to eight weeks for retrieval-heavy assistants like Perplexity, because they read the live web and pick up a new page within days once it is crawlable and cited elsewhere. Assistants leaning on training data move far slower, sometimes only after a model refresh. If your number has not budged in two weeks, that is not evidence the work failed.
Manual checking works for the very first look: open two assistants, ask five buyer questions, see if your name appears. It stops working when you need a trend, because you cannot manually run 30 prompts times 3 samples times 4 models every week and keep the conditions identical. The switch usually happens once you start changing things and want to know whether the changes helped.
Start with reachability rather than content: confirm AI crawlers are not blocked in robots.txt, that your key pages render without JavaScript, and that your product has a definition page an assistant can quote. A brand with no crawlable description and no third-party listings has nothing for a model to retrieve, so no amount of new blog posts will move the number.
Mostly category and comparison prompts, since those are where buyers get a recommendation and where you can lose to a competitor. Keep two or three branded prompts as a hallucination check: asking what your product does will reveal if an assistant has your pricing, positioning, or feature set wrong, which is a different problem with a different fix.
As a working floor, 20 to 30 prompts sampled a few times each. Below that, the score mostly reflects which questions you happened to choose, and week-to-week swings will be noise rather than signal. Adding prompts past roughly 50 tends to sharpen the picture less than adding samples per prompt.
