What have the authoritative sites measured recently?

On 2026-06-15, Semrush published “What is AI sentiment analysis? A marketer’s guide.” It is not about data from any one industry but about a method: how to judge whether AI treats a brand positively, neutrally, or negatively in its answers, and why this deserves monitoring in its own right.

Their reasoning: a single sentence in an AI answer may shape a buyer’s first impression more than ten reviews, because users treat it as a neutral summary rather than advertising. That judgment has data behind it — on 2026-10-01, Ahrefs measured that 83% of branded searches return an AI Overview, and people searching your company name often see that summary first.

What interests me is the other side: this method sounds like it requires tools, but it can be done by hand. Semrush’s article is aimed at teams with a budget; I’ll break it down into a version that costs nothing. The reason is simple: most foreign-trade sites still have low AI mention volumes, so manual work is entirely manageable. For reference definitions, start with how we measure mentions, citations, and traffic.

Mentioned, recommended, described correctly: three different kinds of success

Many people treat “AI mentioned us” as a single outcome. In my panel I split it into three layers, because their difficulty and significance are completely different.

The third layer is the easiest to overlook. I have seen AI describe a company that makes precision parts as a general hardware supplier, and in Chinese answers I have seen service scope described incorrectly. Being mentioned but described wrongly is more troublesome than not being mentioned at all, because it is delivering the wrong positioning to customers on your behalf. I wrote about how to fix it in how to fix it when AI answers get your brand wrong.

So the monitoring table needs at least two columns: which layer an entry belongs to, and whether the description departs from the facts. With only the first column, you will reach overly optimistic conclusions.

One more detail deserves its own record: in the same answer, the model may first put you on the list and then add a caveat, such as “good for small batches but not for very large ones.” Is that sentence positive or negative? I mark it separately as a “conditional recommendation” and copy the exact sentence into the notes. That caveat is often exactly what buyers care about, and looking only at praise or criticism will miss it. Copying the caveat in full is more useful than marking it “positive.”

How do you use the four-tier annotation method?

Fix the question set

Prepare 10 to 15 questions covering three types: direct brand questions (what is your company like), category comparisons (how to choose this kind of supplier), and scenario questions (is your product suitable under a given set of conditions). Once the question set is fixed, do not change it midway, or the rounds will not be comparable.

Four-tier annotation

Give each answer one label: positive, neutral, negative, or not mentioned. Write the criteria down in advance; do not go by feel. My criteria: positive means the answer clearly endorses or recommends; neutral means it merely states facts; negative means it raises a clear drawback or warning; not mentioned means the brand does not appear. Only when the criteria are written down can different people’s labels be compared.

Record models and dates separately

Ask the same question across different models and annotate each separately; do not merge them into one average score. Different models have different training data and retrieval behavior, and mixing them erases the differences. See how AI recommendations take shape to understand how models differ.

Sentiment shifts in our own panel

The T0 baseline round was on 2026-09-17 and T1 on 2026-10-12, with 72 answers each. Strict brand mentions rose from 3/72 to 6/72, and name-drops in service-provider questions rose from 0 to 2. Both sound good, but broken down by the three layers, they mean something completely different at layer two and layer three.

Those 2 name-drops appeared in questions like “how do I choose a service provider” and “why are quotes so far apart,” which counts as being recommended. The earlier 3 mentions mostly just cited our articles as reference material, without recommending our services in the answer. Blending the two into a single “AI visibility improved” conclusion would overstate the progress.

There is another warning in the same round of data: cited domains went from 322 to 245, with an overlap of only 90. The models are switching sources, and a positive description today may not be there tomorrow. That is why you record monthly and keep the original text, rather than taking a screenshot at a single moment. The method is in our testing research and method notes.

What to do first when you find a negative assessment

First work out whether it is true. If it is — the lead times really are long, the certifications really are incomplete — then the job is to change the business, not the wording. AI has simply said out loud a problem that already existed. The right move in that case is to write the improvements up as public information so the next retrieval can read the new situation.

If it is a misjudgment caused by stale information or a confused mix of sources, go back to the sources. The official website comes first, authoritative directories and platform pages second, and social media third. The order cannot be reversed, because AI trusts structured, traceable sources more. I broke down how to unify identity and positioning in brand authority and AI search.

Two things not to do: first, issuing statements everywhere, which only adds noise; second, pretending not to see it, because the answer will keep being wrong along with the old information. The value of monitoring is catching this kind of error early and dealing with it before it affects real inquiries.

A low-cost implementation checklist

Three steps, one round a month, no money spent.

1. Fix 12 questions: four each for brand, category, and scenario; write them down and do not change them again.
2. Build a four-tier table: date, model, question, original text, label, and whether the description is accurate — six columns is enough.
3. Review the negative and “described wrongly” categories once a month: pick one item from each and fix the corresponding source; do not change ten things at once.

After three full months you will have a record that fits your business better than any sentiment tool. To turn it into a routine, see the AI visibility monitoring checklist and how to define your monitoring criteria.

The questions bosses ask most often

Do you have to buy a tool for AI sentiment monitoring?

No. A fixed question set plus four-tier manual annotation is enough to start. Tools save time and scale, but only once you have thought through what you are annotating.

If the same question gets different answers from different AIs, which one is the standard?

Neither. Record each model’s answer and label separately; what you watch is the trend and the consensus, not the wording of any one run.

AI says something bad about us — can we ask it to delete that?

There is no such channel. What you can do is improve the sources it reads: unify your identity, add facts, correct old data, and then keep watching whether the answer changes.

How often should you monitor?

Once a month is enough. AI answers do not change quickly, and frequent re-checks only add noise and tempt you into overreacting to a single fluctuation.