You know ChatGPT names you. Good. But that’s only half the story, and often the less important half. The other half is the tone: when AI talks about your brand, does it recommend you with enthusiasm, mention you in passing, or drop a "but" that kills the sale? Showing up and being liked aren’t the same thing. A brand can appear in nine out of ten answers and lose the sale in all of them because the model describes it as "expensive," "hard to work with," or "worse than the alternative." Here’s how to measure that tone —positive, neutral or negative—, why it matters more than the mention itself, and what to do when it turns against you.
Showing up isn’t being liked: the mention doesn’t tell you the tone
Tracking mentions and auditing your visibility answer the same question: do you show up? Mention tracking counts how many answers you appear in; a GEO audit gives you the snapshot of where you stand today. Both are necessary and both share a blind spot: they tell you that you exist for the model, not how it treats you. And the moment a buyer asks "is [your brand] any good?" or "what are the alternatives to [competitor]?", the tone of the answer decides more than the position.
Think about it from the other side of the screen. To someone who doesn’t know you, it doesn’t matter that you come up first if the phrase attached to you is "tends to be pricier with slow support." A mention with a "but" glued to it erodes trust right when the purchase is decided. That’s why sentiment is the metric that turns visibility into business: it’s not enough to be in the answer, you have to be well in the answer. Counting appearances without watching the tone is like judging a reputation by how many people talk about you, ignoring whether they speak well.
Where AI gets the tone it uses to mention you
The model doesn’t invent the tone: it inherits it from what it has read about you. And it doesn’t read it from your site —where, naturally, everything sounds great— but from the same club of third-party sources AI cites: reviews on G2 or Capterra, Reddit threads, media comparisons, your sector’s forums. If "slow support" dominates in those sources, the model will repeat it even if you fixed the problem a year ago. A negative tone almost never comes out of nowhere: it comes from a public conversation you’re not watching.
There’s a second, subtler source of tone: the comparative frame. When the model puts you next to a competitor, it inherits the dominant narrative of that comparison —who’s "the premium," who’s "the cheap," who’s "the complicated" one—. That pigeonhole isn’t in any single review; it’s the average of how the category is talked about. Changing it is slower than fixing one review, but it starts the same way: knowing what’s being said. You can’t move a tone you haven’t measured.
How to measure sentiment in ChatGPT, Perplexity and Google AI without a paid tool
You don’t need to buy anything to know what face AI mentions you with. You need a battery of questions, a classification rule and a cadence. Three steps:
- The branded battery. Unlike a visibility audit —which avoids your name to measure whether you get in on merit—, here you DO put your brand in. Prepare 8-12 questions a real buyer would ask knowing you exist: "is [your brand] any good?", "how’s [your brand]’s support?", "[your brand] vs [main competitor]", "is [your brand] worth it for [use case]?". These are reputation questions, not presence questions, and that’s why they live in a separate set from the one you use to track mentions.
- Classification into three buckets. Fire each question, in an incognito window, in ChatGPT, Perplexity and Google AI, and classify the answer with a fixed rule —don’t eyeball a score—. Positive: it recommends you or highlights a strength with no caveats. Neutral: it describes you without judging, with flat data. Negative: a "but," an objection or an unfavorable comparison shows up. Also note WHICH exact phrase sets the tone: that quote is what you’ll attack later.
- Cadence and the series. Rerun the same set every two to four weeks and save the date. One day’s sentiment is an anecdote; a three-month series is what tells you whether you’re getting better or worse and whether a model change has changed your face. Without a fixed cadence there’s no measurement, there’s curiosity.
| Sentiment | How it sounds in the AI answer | What it measures |
|---|---|---|
| Positive | "One of the best options for X"; "stands out for its support"; recommends you with no caveats. | Healthy reputation: the model uses you as a reliable reference. |
| Neutral | "A tool that does X"; lists you next to others, with flat data, without judging. | Presence without defense: you exist, but you don’t convince. Clear room to improve. |
| Negative | "Tends to be pricier"; "some users report…"; mentions you with a but or loses the comparison. | Active risk: the mention subtracts instead of adding. Priority one. |
One important thing: don’t inflate the sample or change the rule. With 8-12 questions per engine every two weeks you already see a pattern; what matters isn’t the exact count but that the classification is always the same. If one month you count as "neutral" what you counted as "negative" the month before, your series lies. The discipline of classifying the same way is worth more than the most expensive tool. This exercise fits inside the wider framework of how to measure AI visibility: sentiment is one of the metrics almost nobody watches and one of the most decisive. Measure it alongside presence, not instead of it.
What to do when the tone is negative
Spotting a negative tone without acting is collecting bad news. The response depends on where the "but" is born:
- If the but is a real, current objection (slow support, unjustified price): the work isn’t marketing, it’s product or messaging. Fix the cause or explain the reason for the price better in the sources AI reads, not just on your site.
- If the but is old and you already solved it: the problem is that the public source hasn’t updated. Go to the reviews and forums where that complaint lives and work to make the new narrative outweigh the old one —official replies, recent cases, fresh reviews from happy customers—.
- If the but is a comparative pigeonhole ("the expensive one," "the complicated one"): reframe the comparison with content that puts your strength on the right axis. Don’t fight to be "cheaper" if your edge is something else; change the criterion you’re compared on.
- If all the sentiment drops at once without you touching anything: suspect a model change before a reputational disaster. Wait one measurement cycle before reacting hot.
In all four cases, the pattern is the same: sentiment is fixed off your website, in the sources that shape the model’s opinion. Editing your landing page doesn’t change what AI read on Reddit. That’s why moving the tone is a different job from measuring it, and almost always slower: first the public conversation changes, then what the model repeats changes.
When to stop measuring by hand and delegate it
The spreadsheet holds up fine while you have a handful of questions and a single market. It gets too small for the same three reasons as always: the volume of questions or languages eats your morning; you need automatic history and competitor comparison without transcribing every answer; or you want someone watching the tone continuously to warn you of a narrative shift before it costs you sales. There, delegating AI visibility and sentiment monitoring gives you the watching done, not a pretty dashboard. And when it’s time to move the tone —rewriting the narrative in the sources the model weights—, GEO optimization does the work, not the report. If you’re still starting out, the full map of the cluster lives in the SEO for ChatGPT guide.