Five brands appear in an answer about your category and yours is one of them. Ya tracker records a mention. The sentence containing your name tells the reader to consider you only if they have a niche requirement most of them do not have, and the next paragraph explains why the other four are safer choices. The count says you are present. The text says you are a footnote.
Mention counts are the cheapest AI visibility metric to compute and the easiest to misread. This post separates the ways a brand can appear in an answer, explains why classifying them is harder than it looks, and gives a way to record framing that does not pretend to a precision the method does not have.
Every example below is illustrative. None is a transcript from any assistant, and none should be read as a description of how a specific product phrases anything.
Four ways your name can appear in an answer
A mention is a string match. Four quite different things produce one.
1. Recommended. The answer puts you forward as a choice the reader should make, usually with a reason attached and usually first or near-first when a list is involved. Illustrative form: "For teams under twenty people, [Brand] is usually the simplest place to start, because setup takes minutes and it does not require a dedicated administrator."
2. Listed. Your name appears in an enumeration with little or no differentiating comment. Illustrative form: "Common options include [Brand A], [Brand B], [Brand C] and [Brand D]." This is the largest category in practice and the one most often mistaken for the first. It carries real value, because appearing in the consideration set is a prerequisite for anything else, but it is not an endorsement and should never be counted as one.
3. Compared unfavorably. You are named specifically in order to be positioned below something else. Illustrative form: "[Competitor] covers the same ground as [Brand] and includes reporting that [Brand] charges separately for, so most buyers comparing the two end up with [Competitor]." Your mention count goes up. Your commercial position goes down.
4. Cautioned against. The answer names you attached to a risk, a limitation, a complaint pattern, or a warning. Illustrative form: "Buyers on smaller plans have reported that support response times with [Brand] can be slow, so if support matters to you, check the terms before committing."
Categories three and four are the reason a rising mention chart can accompany a flat pipeline. A metric that treats all four as identical will report growth during a reputational decline.
Recommended, listed, compared unfavorably, cautioned against
The four are not a neat scale, and treating them as one number from bad to good hides the useful distinctions. They differ on two axes worth tracking separately.
| Appearance type | Are you a candidate? | Does the framing help or hurt? | What it usually means |
|---|---|---|---|
| Recommended | Yes, foregrounded | Helps | Present in the sources the system leaned on, and described in terms that match the question asked |
| Listed | Yes, undifferentiated | Neutral | Known as a category member; nothing in the retrieved material distinguishes you for this question |
| Compared unfavorably | Yes, as a reference point | Hurts | Comparison content exists and does not favor you, or a competitor's positioning material dominates the comparison |
| Cautioned against | Sometimes not | Hurts, sometimes badly | A negative theme has enough presence in the sources to survive into a synthesized answer |
Two practical readings follow. Moving from listed to recommended is a differentiation problem: the answer knows you exist and has nothing specific to say about you. Moving out of cautioned against is a different job entirely, closer to a correction and reputation problem, and worth handling before any work aimed at the first.
There is also a fifth case worth logging even though it is not a mention: the answer describes something you do, in your words, without naming you. That is a real signal about how retrieved material is being used, and a pure string-match metric will never see it.
Why sentiment classification is harder here than in social listening
Social listening tools classify short, opinionated, single-voiced text. A tweet has one author with one attitude. Assistant answers break most of those assumptions.
Length and mixed valence. A single answer can recommend you for one use case, warn against you for another, and stay neutral in a comparison table, all in four hundred words. There is no single sentiment for the document. Assigning one requires a rule about which part counts, and that rule is a judgement call you should write down rather than inherit silently from a tool.
Conditional framing. "If you need X, [Brand] is the obvious choice" is positive only if the reader needs X. The same sentence is an exclusion for everyone else. Sentiment models trained on declarative opinion text handle conditionals poorly.
Hedged and attributed statements. "Some users report" and "reviews suggest" carry attitude at one remove. Whether that is negative sentiment about you or neutral reporting about reviews is genuinely ambiguous, and different annotators will disagree.
Position and structure matter more than words. Being third in a list of three is a weaker outcome than being first, with identical wording. Sentiment classifiers look at words, not rank order, so they miss the strongest available signal.
Comparative sentences have two subjects. "[Competitor] is faster than [Brand]" is positive for one name and negative for the other in the same clause. Whole-document or even whole-sentence classification will get one of them wrong.
The generated text is not evidence about the world. A caution appearing in an answer means the theme was present in what the system drew on. It does not establish that the underlying claim is accurate, and if it is inaccurate, that is a correction matter rather than a measurement one.
Anyone who tells you sentiment on this material is a solved classification problem is selling something. We do not know the accuracy of automated framing classification on long assistant answers, because it depends on the classifier, the category, the answer length, and the labelling scheme, and there is no settled public benchmark for it that we would stake a number on. What we can say is that error concentrates exactly where the stakes are highest: mixed-valence answers, conditionals, and comparative clauses. Those are the cases you most need to be right about.
How to record framing without pretending it's exact
The goal is a record that is useful, reproducible, and honest about its own softness.
Use a coarse scheme. Four labels, or five with the unnamed-but-described case. Finer-grained scoring on this material implies a precision the underlying judgement does not support. A ten-point sentiment score on an assistant answer is false precision.
Classify at the mention, not the document. One answer can contain two mentions of you with different framings. Code each occurrence, then summarize the document as a distribution rather than a single label.
Write the codebook before the first run. Define each label with two or three real decision rules, including how to handle conditionals, hedged attribution, and comparative clauses. Version the codebook the way you version the prompt set, and attach the version to every result. Recoding to a new scheme resets your trend line.
Record position alongside framing. Order of appearance in a list is cheap to capture and often more informative than the framing label.
Report the machine label and the human check separately. If an automated pass produces the first cut, keep its output as its own field rather than overwriting it with the human decision. The disagreement rate between the two is your running estimate of how much the automated number can be trusted, and it is specific to your category rather than borrowed from someone else's benchmark.
Spot-check on a stated cadence. A workable default: a human reviews a random sample of at least 10% of coded mentions every reporting cycle, with a minimum floor of twenty mentions per cycle so a small run still gets meaningful review, and a full manual pass on any prompt where the framing label changed from the previous run. State the cadence in the report. If a cycle got skipped, say that too.
Keep the raw text. Store the full answer with every coded mention. Every framing dispute is resolved by reading the answer, and no dispute is resolvable if you only kept the label.
Which of these fields belongs in a headline number is a separate decision, covered in which metric answers which question. Framing is also distinct from citation: a link to your site and a favorable sentence about you are different events, and what counts as a citation draws that line.
What to report when the framing is mixed
Mixed is the normal case, not the exception. Report it as a distribution and resist collapsing it.
A defensible reporting line looks like this:
Report the total answers containing the brand, run dates, model versions, prompt-set version, and codebook version. Then state the counts coded as recommended, listed, compared unfavorably, and cautioned against, plus the share reviewed by a human and the disagreement count.
Four rules go with it:
- Never publish a single "sentiment score" as the headline. Publish the distribution. If a summary figure is unavoidable, define it explicitly and show the distribution next to it.
- Report the negative categories as counts, not percentages, when volume is low. Three cautions out of eleven mentions is not "27% negative" in any stable sense.
- Separate framing change from volume change. Mentions can rise while the recommended share falls. Both belong in the report, and the second is usually the more important one.
- Flag any caution mention containing a factual claim about your product for review outside the measurement process. If the claim is wrong, the response is a correction workflow, not a metric.
The uncomfortable version of this report is the useful one. A quarter where mentions grew and the recommended share shrank is a real finding, and a team that surfaces it early has time to act on it. A team that reports the mention count alone finds out from the pipeline six months later.
In Plain English
Being named in an AI answer covers four different situations: recommended, listed, compared unfavorably, and cautioned against. Only the first is an endorsement, and mention counts treat all four as the same event. Classifying which is which is a judgement call, not a solved technical problem, so record it with a coarse scheme, a written codebook, and regular human spot-checks rather than a single automated score.
This is for anyone whose mention chart is going up while nothing downstream moves: growth leads, agency strategists reporting to clients, and brand and comms teams who need to know when a negative theme starts surviving into answers. Do this work when you first set up tracking, and re-read the raw answers whenever a number moves more than you can explain.
Next step: pull twenty recent answers containing your name, code each mention by hand into the four categories, and compare the result with what your dashboard reported for the same period. The gap between the two is the size of your interpretation problem.
Run the free AI Visibility Check to see how your brand is framed in sampled answers, not just how often it appears. Details of the sampling and coding approach are on how to measure AI visibility, and you can create a manual snapshot against your own category.
Primary-source verification
Checked July 28, 2026. The operational and factual guidance above was reviewed against the primary references below. Provider features, prices, policies, interfaces, jurisdictions, and account-specific results can change; recheck the controlling first-party document and your own account or measured data before acting.
- Google Search: AI features and your website - How Google describes eligibility, preview controls, and appearance in its AI search features.
- Google: common crawlers and special-case user agents - Current Googlebot and Google-Extended identities and controls.
- OpenAI: overview of OpenAI crawlers - Current GPTBot, OAI-SearchBot, and ChatGPT-User descriptions and publisher controls.
- Anthropic: web-crawler controls - Current ClaudeBot, Claude-SearchBot, and Claude-User descriptions and controls.
- Perplexity: crawler and user-agent documentation - Current PerplexityBot and Perplexity-User descriptions and controls.
- FTC: Endorsement Guides Q&A - Current U.S. regulator guidance on endorsements, disclosures, and reviews.
- FTC: Consumer Reviews and Testimonials Rule Q&A - Current U.S. rule guidance on deceptive review practices.