The short answer
Track what an answer says about your brand as well as whether it names you. Separate recommendation, positioning and factual accuracy. Keep the exact quote, the buyer question and the supporting source. Correct consequential errors with clear evidence; a flattering description can still put you in the wrong category.
“An excellent budget option for small teams.” Lovely sentence. Slight problem if you sell enterprise software with a six-month procurement cycle.
A mention counter records a win. A positive-sentiment label probably agrees. Your sales team would like a word. This is why a brand audit needs a little more reading and a little less celebration of green numbers.
Three checks that a happy face cannot do
| What the answer says | What to check | Why it matters |
|---|---|---|
| “A good option for freelancers” | Positioning: audience and buying job | May attract the wrong buyer even when positive. |
| “Does not offer single sign-on” | Accuracy: feature, plan and date | A verifiable mistake could remove you from a shortlist. |
| “Strong reporting, but complex setup” | Trade-off: evidence and context | A fair limitation may help buyers choose appropriately. |
Keep the checks separate. Sentiment describes tone. Positioning describes who the answer thinks you serve. Accuracy concerns claims you can verify. “Expensive” may be a reasonable judgement; “starts at CHF 900 per month” can be checked against a dated price page. Treating both as negative sentiment produces a bad to-do list.
A recommendation also depends on the question. A complex implementation may be acceptable for an enterprise buyer and a deal-breaker for a two-person team. Preserve the prompt beside the description. Otherwise the quote loses the very context that made it useful.
Make a sheet someone else can check
Start with answers in which your brand actually appears. For each distinct claim, copy the wording without tidying it up. Record the provider, date, question, cited URL where available, and the official page you used to check it. One answer containing the same claim three times is still one answer.
- Use “unclear” when the inspected page does not settle the question. Missing evidence is not proof that a claim is false.
- Count the answers containing a claim, then show the denominator: “incorrect in 4 of 18 answers mentioning us”, alongside the total sample size.
- Keep claims about a similarly named company separate. An identity mix-up needs an identity fix, not a warmer adjective.
- Ask a second person to review disputed labels. Store the disagreement if it cannot be resolved.
The output should survive a colleague asking “where did that come from?” A single average sentiment score cannot answer that question. A short claim ledger can.
Fix the misunderstanding, not the customer
Prioritise errors that could change a purchase: availability, supported market, key integration, price, security claim or target audience. Before adding a page, check whether a good explanation already exists. The problem may be a stale comparison page or an ambiguous product name rather than a content gap.
| Observed problem | Reasonable next step | Evidence to keep |
|---|---|---|
| Old price repeated | Clarify the current price and effective date; request correction on a cited stale source. | Old quote, current price page, publication date. |
| Wrong audience | Make the use-case page explicit about who it serves, with a relevant documented example. | Buyer question, misleading wording, inspected destination. |
| Fair product limitation | Explain the trade-off accurately where buyers evaluate it. | Verified limitation and the audience for whom it matters. |
Do not invent reviews, conceal real limitations or promise an integration the product does not have. And do not turn a marketing report into a product-roadmap order. Some answers will correctly recommend a competitor. That is information, not a malfunction.
Check whether the same mistake comes back
Save the date of the correction, allow for discovery, and rerun the same questions with the same recorded settings. Compare the error counts by provider. Keep unaffected questions in the panel too: if everything shifts together, a model or retrieval change may be involved.
A disappearing error is encouraging, but one before-and-after run does not establish why it disappeared. Our guide to score reliability explains how repeats help you avoid mistaking ordinary movement for a result.
The useful ending is specific: “The old price still appears in two saved answers; we corrected our pricing page and asked the cited directory to update its entry.” That gives somebody a job. “Improve sentiment by 15%” gives somebody a headache.
Sources & context
This guide responds to a question raised in In AI visibility, are you tracking just Share of Voice or also sentiment? on Reddit. The discussion informed the topic; it is not evidence of the effectiveness of any product mentioned there. Platform documentation is linked alongside the claims it supports.
Examples and figures are illustrative unless explicitly stated otherwise. Written and published by Share of Voice, a Superstellar LLC product. We sell AI visibility reports; our commercial perspective is worth keeping in mind.
Spotted something we should correct? Let us know.