AI Brand Sentiment Analysis: How Many Runs You Need

AI brand sentiment analysis measures the tone an answer engine uses when it describes, compares or recommends your brand. An answer engine is a service such as ChatGPT, Perplexity, Gemini or Google AI Mode that writes a direct answer instead of a list of links. Social listening measures something else: what people themselves post.

The two are linked, because engines cite community sites. Our research at Citation Research, covered in the triage section below, found Google AI Mode citing Reddit, YouTube or Quora in most answers. But the engine compresses all of that into a few sentences, and a buyer reads only those sentences. So you log four things separately:

• Mention: whether your brand appears at all.
• Tone: positive, neutral, negative or mixed.
• Aspect: what the tone is about, such as price, setup, support or reliability.
• Accuracy: whether any checkable claim about you is true.

Three terms are used the same way throughout. A prompt is a question a buyer would type. A run is one answer to one prompt on one engine, asked in a fresh conversation. A reading is the percentage you compute from many runs.

A small sample of AI sentiment answers misleads because tone flips between positive and negative far more often than whether a brand is mentioned. In the June 2026 Ranqo preprint, tone flipped in 45.5% of qualifying combinations and mentions in 6.8%, so tone is about 6.7 times noisier.

The paper analyzed more than 100,000 responses across more than 100 brands tracked on Ranqo's platform from March to May 2026. Its unit is a brand, prompt and engine combination. It recommends at least 10 distinct prompts per engine per brand before a sentiment-weighted score stabilizes.

Read it with care. The author's company sells AI visibility tracking, the brands skew toward software, retail and fintech, and the paper says its sentiment labels come from a model, so the 45.5% mixes real answer variance with classifier error. The direction is still clear. The same paper found no combination that was consistently negative, which suggests negativity tends to be intermittent. That is our inference, not a finding about your brand.

Practitioners see the same instability across tools. "If you use three different tools and give them the same prompts, you get three different answers," Paul Dyer, CEO of /prompt, told Digiday in May 2026. In one test by Marketing With Dave of six tools on a single branded prompt ("Who is Marketing With Dave?"), visibility on one engine came back as 100%, 90%, 75% and 16.7%. That is one prompt and one small brand, so treat it as an anecdote that shows the problem, not a measure of it.

An AI sentiment reading needs at least 10 answers per engine before you show a percentage, and about 30 before you compare it with another period. Fewer answers give a range so wide that 60% positive and 40% positive cannot be told apart.

The table is our calculation. For each row we took the positive runs out of the total and computed a 95% Wilson score interval, a standard range for a share measured on a small sample. Read the last column as: the true positive rate could plausibly sit anywhere in this span.

For a range of plus or minus 10 points, the standard formula n = z² × p × (1 − p) ÷ d² gives about 81 runs when the true share is 70% (z = 1.96, d = 0.10), about 96 at 50% and about 35 at 90%. These assume each run is independent, which fresh conversations approximate but do not guarantee.

The preprint recommends at least 10 distinct prompts per engine per brand. Our own rule, based on the interval table, is at least 10 answers per engine before you show a percentage, whether they come from one prompt repeated or from several prompts pooled. Pooling means combining runs from different prompts into one reading. A pooled reading describes your prompt set on that engine, not any single prompt, so do not mix pooled and repeated-prompt readings in one number. This is our recommendation.

Log every run with the same columns and label it with fixed rules, so two people reading the same answer reach the same label. Copy the columns into a spreadsheet.

Labeling rules, each of which you can test on a sample:

• Positive: the engine recommends you, praises you with a stated reason, or lists you among its top picks.
• Negative: the engine warns against you, gives a drawback as a reason to avoid you, or ranks you below a named alternative with a reason.
• Neutral: the engine describes you without recommending or criticizing.
• Mixed: both a recommendation and a criticism appear. Do not force these into positive or negative.
• Wrong fact: a checkable claim, such as price, feature, location or founder, that contradicts your own pages. Flag it separately from tone.

Agreement test (our recommendation): have two people label the same 20 answers independently. Pass if they agree on at least 18. If they do not, tighten the rules before you collect more data, because a vague rubric adds noise on top of the engine's own.

Collection rules: use a fresh conversation for every run, because earlier messages shape later answers. Spread runs across several days, keep the country constant, and note the engine and model version if shown. By hand, 20 prompts on 3 engines at 10 runs each is 600 answers to label.

A change in AI sentiment is real only when the 95% ranges of the two periods do not overlap. Overlapping ranges mean the data cannot distinguish the two periods, whatever the headline numbers say.

A worked example, with our calculation, on one engine. Last month: 21 positive out of 30 runs, 70%, range 52% to 83%. This month: 15 positive out of 30, 50%, range 33% to 67%. That looks like a 20-point drop, but the ranges overlap between 52% and 67%, so you cannot call it a decline. With 100 runs each, the same shares give 60% to 78% and 40% to 60%, which do not overlap, and the drop is real.

Two things can still fake a change. An engine update can shift tone for every brand at once, so check whether competitors moved the same way. A change in the prompt set, the country or the labeling rubric makes two periods incomparable. Keep all three fixed between periods.

Triage a bad reading by cause before you write anything. Check for noise first, then wrong facts, then the source the engine drew on.

Do the wrong-fact check even when your tone is fine. In the same Marketing With Dave test, a tool reporting 90% visibility for the brand attributed it to the wrong person. A high mention rate can hide a false claim.

Engines also differ in which sources they lean on. In our research at Citation Research, Study 007, we asked 150 questions of each of four engines from India on October 5, 2026. Google AI Mode cited a community platform (Reddit, YouTube or Quora) in 117 answers, Perplexity in 73, Gemini in 29 and ChatGPT in 3. The study did not measure sentiment, so the conclusion is our inference: a negative forum thread is more likely to shape the tone in Google AI Mode than in ChatGPT, and you should read each engine separately. For how mentions move over time, see our guide to tracking AI citations over time.

PageLens.ai tracks a fixed set of buyer-intent prompts on a schedule across the answer engines your plan covers, and applies sentiment analysis to the answers on every plan. Verbatim answer evidence is also included on every plan, so you can read the exact sentence behind any label and run the wrong-fact check above.

Plans as of October 2026, in US dollars, from our pricing page:

Sentiment analysis and verbatim answer evidence are included on the Agency plan too.

The answer counts equal tracked prompts times engines: 100 × 3 = 300, 100 × 5 = 500 and 200 × 7 = 1,400. So each plan collects one answer per prompt per engine per collection period, weekly on Launch and daily on Growth and Enterprise. A 30-run reading for one prompt on one engine therefore takes about 30 days on Growth or Enterprise, and about 30 weeks on Launch. To reach 30 runs sooner, pool prompts: 100 prompts on one engine give 100 runs in a single day on Growth, and in a single week on Launch. Only answers that mention your brand carry a tone, so the usable count is smaller than the answer count.

Our recommendation follows from the run table. Report monthly, pool your prompt set per engine for the headline reading, and keep a few core prompts running long enough to reach 30 runs each. Before you buy any tracker, our AI visibility tool trial checklist lists what to test over 14 days. To see sentiment on your own prompts, book a demo.