Before you read the results
Your AI visibility results are a record of what several assistants said to a fixed set of questions, sorted into four outcomes: absent, mentioned, recommended, or cited with a link. Reading them well means knowing what each outcome means and what it does not.
This page assumes you already have results, from your own manual check or from software. If not, our page on producing a set of results explains how to get them. Everything here applies whichever way you collected them.
Two cautions before the detail. First, every figure is a sample. The same question can give different answers on different runs, so a result for one question on one day is a hint, not a verdict. Our page on why AI answers change explains why, and how many runs make a fair reading. Second, results are only comparable if the questions, assistants and setup stay the same, so check that they did before reading any movement as news.
With that said, here is how to read what you have.
The four outcomes
For each answer in your set, you can place your business in exactly one of four outcomes. They form a ladder.
Absent
Nothing in the answer mentions your business, links to it, or alludes to it. The assistant produced a complete reply without you. This is the baseline state for any business on a question it has not yet earned.
Mentioned
Your business is named, but the assistant does not push it. It might be one of six names in a list, or a passing reference. The assistant knows you exist in this category, and that is all it has said.
Recommended
The assistant puts you forward. You are first in a shortlist, described as a good fit for the situation, or named as the answer when it picks one. This is where results start to resemble sales.
Cited with a link
The answer links to your website, or clearly names your page as a source. This is a statement about where the assistant got its information, and it is separate from whether you were recommended. It sits at the top because it shows your own pages are doing the work, and because the reader can click straight through.
These outcomes can combine. A business can be recommended and cited, which is the strongest case, or mentioned and not cited, which is common. When you record results, note recommendation and citation separately, since they respond to different causes.
Reading an absent result
Absent is the most common outcome, and the least informative on its own. You need to ask why you are absent, and the answer depends on what else is in the answer.
If the answer names a handful of businesses and cites sources, look at those sources. Are they review sites, local guides, forums, comparison pages? If you are not on them, that is the reason. The assistant read those pages and you were not there.
If the answer is generic, with no names, the question may be too broad to produce businesses at all. That is worth knowing, because it tells you the question is not where buyers choose.
If you are absent from a question about your own core service, in your own area, it usually means the assistant has too little to go on. Our page on why a brand is not in ChatGPT covers the common reasons.
One more point. A result of absent in one run, with presence in other runs of the same question, is not absent at all. It is borderline. Count the share across runs before filing it under absent.
Reading a mentioned result
Mentioned is a half-win. You are known, and you have not been chosen.
Look first at where in the answer you appear. Named in the first sentence of a short list is stronger than named seventh in a long one. Then look at the wording. If the assistant says nothing more than your name, the mention is thin. If it adds a detail, even a small one, there is substance to build on. Our page on brand visibility in AI search goes through how to read that wording.
Then ask what separates the recommended businesses from you. Often it is one clear thing: a page that compares options, a stronger set of reviews, a specific statement of who they are for. That one difference is your homework.
Mentioned results are also the easiest to improve, since the assistant already has you in mind. A lot of the quickest gains come from turning a mention into a recommendation by giving the assistant something specific to say.
Reading a recommended result
A recommended result is good news, with three things to check before you celebrate.
How often? One recommendation in one run is a draw. Look at the share across runs. If you are recommended in four runs of five, that is a reliable result. If in one of five, you are borderline.
For what? The reason matters. Recommended as the cheapest, or the closest, or the fastest is a narrower win than recommended as the best fit. Check the reason against what you want to be known for.
Is it accurate? A recommendation based on a wrong fact will fail the buyer and, over time, you. If the assistant says you offer something you do not, fix the source of that claim.
Notice, too, whether you were recommended without being cited. That means the assistant has a view of you that is not coming from your own pages. It may come from reviews, directories or press. That is fine, but it is outside your direct control, and it is worth finding out whose pages are carrying you, since you depend on them.
Reading a cited result
Being cited with a link means your own page was used or shown as a source. For many businesses this is the outcome they most want, because it brings both credit and a path to your site.
Check which page was cited. If it is the page you would have chosen, the assistant found what you wanted it to find. If it is an odd page, such as an old blog post or a thin category page, you have learned which of your pages the assistant prefers, and perhaps that your best page is not being picked up.
Check what the answer says about the page. A citation attached to a sentence that praises you is a different thing from a citation attached to a neutral fact.
Check whether the cited page is the one that actually answers the question. A page that is cited, but too vague to answer, may lose the citation as soon as a better page appears. Write the answer more plainly and more completely on that page.
Finally, note that not all assistants show links in all answers, and some answer without any sources. A low citation count in an assistant that rarely shows links is not a failure. Compare like with like, assistant by assistant.
Reading the spread across your whole set
Individual outcomes matter, but the pattern across the whole set is what you act on. Three views are worth taking.
By question group. Sort your questions into broad, specific, comparison and name-based groups. You may find you are strong on name questions, which only prove people can find you if they already know you, and absent on broad questions, where new customers come from. Or the reverse. The group where you are weakest and the stakes are highest is your priority.
By assistant. Strong in one, absent in another, is common, and useful. It points to different sources, and to which assistant your buyers use. Our page on how to measure AI visibility explains how to report shares per assistant.
By outcome mix. Add up how many answers fall into each of the four outcomes. A set that is mostly absent with a few mentions says one thing. A set that is mostly mentioned with a few recommendations says another. The mix tells you which rung to work on.
Draw the three views on one page if you can. Seeing them together stops you from fixating on a single striking answer.
Traps when reading results
A few ways people fool themselves.
Reading a single answer as the result. Look at the share across runs.
Counting a mention as a recommendation. Keep the definitions fixed and written down.
Celebrating a name-based win. Questions that include your name mostly test whether the assistant knows you. Report them separately from the questions a stranger would ask.
Ignoring who else appears. Your results are relative. If the assistants name five businesses and you are one, that is a different position from being one of twenty.
Moving the goalposts. If you add easy questions after seeing poor results, the new figure is not comparable with the old one. Add them as a separate series.
Stopping at the table. Results are only useful if they lead somewhere. When you have read them, the page on what to do after a check matches next steps to what you found.
A worked example with invented numbers
Here is a small set of results, made up to show the reading. A florist tests fifteen questions across two assistants, three runs each, which gives ninety answers.
Sorted into outcomes, the florist finds fifty-four answers absent, twenty-four mentioned, nine recommended and three cited with a link. By group, the name-based questions account for nearly all the recommendations and citations. The broad questions, such as who does good wedding flowers in town, account for most of the absences.
What does that say? The assistants know the florist exists, since they handle name questions well. They rarely volunteer the florist to a stranger. The three citations all point to one page, an old gallery post. The mentions on broad questions all come with vague wording. And two competitors, named in most broad answers, appear to be carried by a local wedding guide and a review site.
The reading, then, is this. The florist is on the middle rung for being known and on the bottom rung for being chosen. The most valuable work is not more name recognition. It is to get specific, plain pages about wedding flowers on the florist's own site, and to find out whether the wedding guide would feature the shop. And the one page that earns citations deserves a look, since it shows the kind of page the assistants prefer.
Nothing here required special software, only a table and a close read.
What each outcome is worth
People ask how much each outcome is worth in money. Nobody can say precisely, and you should distrust anyone who claims to. The path from an answer to a sale usually runs through a later search for your name or a direct visit, which carries no label.
What you can say is directional. Absent is worth nothing for that question, since the buyer never meets you. Mentioned is worth a little, because it plants your name. Recommended is worth the most in terms of choice, because the assistant has done the persuading. Cited with a link adds a route to your site and a sign your own content is trusted.
A fair way to use this is to treat each tier up as progress and to watch the shift from absent to mentioned and from mentioned to recommended, rather than to attach invented values to them. If you want a single summary figure, our page on the AI visibility score covers what such a score can and cannot carry.
Writing your results up in one paragraph
If you cannot summarize results in a paragraph, you have not understood them yet. Try to fill in this template using your own words and numbers.
"We tested a set number of questions, a set number of times, in these assistants, on this date. We were absent from this share of answers, mentioned in this share, recommended in this share and cited with a link in this share. Our strongest group of questions was this, and our weakest was that. The businesses named most often were these, and the sources the answers leaned on were of these types. The most useful single step is this."
Filling it in forces you to commit to a reading. It also gives you something to compare next time. If the paragraph in three months has the same shape but better numbers, you are moving. If it has the same numbers, something needs to change.
Where RunOctopus comes in
RunOctopus does the work that moves questions up these tiers, and it gives paying customers reports on where ChatGPT, Perplexity, Claude and Gemini name the business. It does not offer a free AI visibility checker, so the free way to produce results like these is the manual method described on this site.
As a service, it is done for you. Ollie, the company's AI, researches the business and then writes and publishes pages on the business's own website, with a team that plans, writes, reviews and publishes alongside it. It works with Shopify, WordPress, WooCommerce and Wix, and the owner has nothing to install, approve, upload or schedule. The first week brings eight long-form articles, six pages and a custom interactive tool for the niche, followed by 175 or more pieces of content a month.
If you want to see Ollie's plan for your site, the free site read at app.runoctopus.com/start reads your address and shows it. It does not test AI answers. No account and no call.