Skip to content

AI answers name your brand. How do you measure whether they recommend you?

A mention count treats a warning and a recommendation alike. How to score the way AI answers describe your brand: strength, caveats, errors and sources.

By Published 14 min read
A figure holds a gauge against an AI answer card, weighing a highlighted brand line between two speech bubbles.

The short answer

Measure whether AI answers recommend your brand by coding every answer that names you on five fields: recommendation strength, the reason given, any caveat, factual errors and the source behind them. Score each engine separately on a fixed prompt set, repeat runs on different days, compare named rivals, and report recommended share beside plain mention share.

Key takeaways

  • A mention count treats an answer that recommends you and an answer that warns buyers off you as the same result, so it needs a second measure on top.
  • In G2's survey of 1,076 B2B decision makers in March 2026, 64% said they encounter inaccurate AI recommendations often or very often, so buyers read the framing closely.
  • A five-level recommendation scale, coded by a person on a sample and by a model on the rest, gives a framing measure that two people can repeat and agree on.
  • Every caveat an engine attaches to your brand should be traced to a phrase on a page, because the fix is almost always on that page and not in the prompt.
  • Our view: report recommended share per engine against named rivals each quarter, and treat a single negative answer as a lead to trace, not as a trend.

Why does a mention count miss what buyers actually read?

Because a mention is only the fact that your name appears. The buyer reads the sentence around it. "Contoso Growth is the usual pick for mid-sized teams" and "Contoso Growth is an option, though some users report slow support" both count as one mention. They do very different work on a shortlist.

Most AI visibility reports stop at the count. Share of answer, the share of answers to your buyer prompts that name you, is the right headline, and our note on mentions and citations as KPIs explains why it beats citations as the number on top. It still tells you nothing about tone, rank in the list or the reason given. A brand can hold its mention share while the engines quietly move it from first choice to fallback.

This method guide adds the second layer: how to code what the answer says about you, so the result can be repeated, compared with rivals and traced to a source you can fix.

Buyers read the framing, and they check it

The framing matters because buyers act on it. In G2's survey of 1,076 B2B decision makers in North America, EMEA and APAC in March 2026, 51% said they start software research in an AI chatbot more often than in Google.1 In the same survey, 64% said they encounter inaccurate AI recommendations often or very often.2

Checking is common too. In TrustRadius's survey of 1,862 technology buyers in January 2026, 63% used AI somewhere in the purchase journey, and 94% of those fact-checked what it told them.3 The same survey found 83% of buyers shortlisted three or fewer products.3 On a list that short, being named with a caveat can cost the place.

Why framing matters

Buyers start in AI, doubt it and shortlist few

51%of B2B decision makers start software research in an AI chatbot more often than in GoogleG2, Mar 2026 1
64%encounter inaccurate AI recommendations often or very oftenG2, Mar 2026 2
83%of technology buyers shortlisted three or fewer productsTrustRadius, Jan 2026 3
Sources: G2, 2026 AI Search Insight Report; G2, inaccurate recommendations; TrustRadius, Beyond the hype.

Prior work and a disclosure

Others have written about sentiment in AI answers. Parse, which sells AI answer monitoring, has published guidance on why being cited is not enough and how to monitor the sentiment of answers that name a brand.4 MaxAEO, which sells AI visibility measurement, has written on sample size and on why repeat runs of one prompt behave as clusters.5 What this piece adds is a coding scheme with fixed levels, a rule for when a shift is real, and a way to trace each caveat back to a page.

Sigzen AI sells AI visibility audits that include a sentiment register, so we have a stake in how framing is measured. Everything below works in a spreadsheet with no vendor at all.

Five fields to record for every answer that names you

Record the same five fields for every answer that mentions your brand, and keep the raw answer text beside them. Fixed fields are what make the measure repeatable. A free-text note such as "a bit negative" cannot be counted, compared or handed to a second reviewer.

The five fields, in the order a reviewer fills them

  1. Recommendation strength

    Where the answer places you, on a fixed five-level scale from first choice to advised against.

  2. Reason given

    The attribute the answer ties to your name: price, specialism, size, support, location of service.

  3. Caveat

    Any limit or warning attached to you, quoted word for word.

  4. Factual errors

    Each claim about you that is wrong or out of date, with the correct fact.

  5. Source

    The cited page, or the page whose wording the caveat or error most likely came from.

Illustrative. The coding scheme the Sigzen AI editorial team recommends; field names can change, the fixed levels should not.

The reason field is the one most reports skip, and it is often the most useful. If every engine recommends you "for small teams" while you sell to large ones, the answer is positive and still wrong for your pipeline. That is a positioning problem the mention count will never surface.

Keep the caveat as a quote. Paraphrase loses the exact phrase, and the exact phrase is what you search for on your own site, review profiles and directory listings when you trace where it came from.

How do you score recommendation strength so two people agree?

Use a short scale with written definitions, and code position and wording together. Five levels is enough to see movement and few enough that two reviewers land on the same level most of the time. Polarity alone, positive or negative, is too coarse: many answers that name a brand read as neutral, and the shift that matters is between "one of several" and "the one to pick".

LevelWhat the answer doesTypical wording
Top pickNames you first or as the single best fit"For this need, start with Contoso Growth"
RecommendedRecommends you among a short group"Good options include Contoso Growth and Fabrikam"
ListedNames you with no judgement"Providers include Contoso Growth, Fabrikam and Tailspin"
QualifiedNames you with a caveat or a narrow use"Contoso Growth suits small budgets, but support is slower"
Advised againstSteers the reader away from you"Users report billing problems with Contoso Growth"

Illustrative. House-roster brands; the wording shows each level, not real answers.

Write the level definitions on one page and give reviewers ten worked answers before they start. Then have two people code the same small sample blind and compare. Where they disagree, tighten the definition rather than averaging the scores. Our view: if two careful reviewers cannot agree on a level, the level is badly defined, and a model coding at scale will inherit the same confusion.

Position in a list is part of the level

An engine that lists five providers usually does not rank them explicitly, but readers treat order as rank. Record your position in any list as a separate number, and let the level reflect explicit wording. That keeps "named first in a neutral list" apart from "recommended as the best fit", which a buyer reads very differently.

Code by hand, by model, or both

Code a sample by hand and the rest by model, and check the model against the people every cycle. Hand coding alone does not scale past a few hundred answers a month. Model coding alone drifts in ways nobody notices, and it is weakest exactly where the measure matters: hedged sentences with a "but" in the middle.

Which coding method fits which job

Volume of answers →

Model onlyFast and cheap, but nobody knows when it starts misreading hedged answers
Model, checked by peopleModel codes every answer; people code a fixed sample each cycle and the agreement rate is reported
Hand onlyAccurate on a small set, slow, and hard to repeat month after month
Two people, blindThe way to write and test the definitions before any model touches them

Need for agreement you can show →

Illustrative. The Sigzen AI editorial team's judgement of when each method holds up.

When a model does the coding, give it the same definitions and worked examples the people used, and ask for the level, the quoted caveat and the reason as separate fields. Store its output next to the human codes for the sample. Report how often the two agree on the level. If agreement drops after a model or prompt change on the coding side, freeze the old coder until the gap is understood.

Keep the coder separate from the engines being measured. A model asked to judge answers written by the same model family may read its own phrasing generously. Using a different family for coding is a cheap way to remove that doubt.

How many answers do you need before a framing shift is real?

More than for a mention count, because framing is coded only on the answers that name you. If you are named in a fifth of answers, a set of a few hundred answers yields a few dozen coded mentions per engine. A small shift in that group can be noise. Report recommended share with an interval, and treat a change as real only when the new interval clears the old one.

Repeat runs do not count as fresh evidence. MaxAEO's guidance explains why answers to the same prompt cluster, so running one prompt many times on one day adds less information than running many prompts on different days.5 Spread runs across days, count prompts as the unit, and our note on how often to re-run checks sets out a cadence that balances cost against noise.

Weight by prompt, not by run. If one prompt produced ten answers and all ten called you expensive, that is one finding about one prompt, repeated. Our view: a single strongly negative answer is worth tracing the same week, but it is a lead, not a trend, until it shows up across prompts and days.

A figure climbs three steps that rise from a plain name tag to a speech bubble and then to a star at the top.
Most answers read alike; the few that differ are worth lifting out and examining.

Where does a caveat come from, and how do you trace it?

Trace a caveat by searching for its exact phrase, first in the sources the engine cited and then across the pages it most likely learned from. Most caveats are not invented. They echo a review, a forum thread, an old pricing page or a rival's comparison page that describes you in those words.

Much of what an engine says does not come from a live search at all. In Ahrefs' study of 66.6 million free, logged-out ChatGPT responses in August and September 2026, across all countries, ChatGPT searched the web in 14.4% of them.6 The rest came from what the model already held. So an answer with no citations can still carry a caveat that traces back to a page written months ago.

Tracing one caveat to a fix

  1. Quote the caveat
  2. Check cited sources
  3. Search the exact phrase
  4. Find the page that says it
  5. Fix or answer that page
  6. Re-run the prompts
Illustrative. The tracing order the Sigzen AI editorial team uses; the highlighted step finds most sources.

Your own pages are often the source. Profound, which sells AI visibility software, published an analysis of inaccurate claims checked by its own product and found pricing and billing to be the topic with the highest inaccuracy rate.7 It traced wrong claims to brands' own pages as well as to third-party and competitor sites.7 Its sample is its customers, so treat the ranking as indicative.

When the caveat is a plain factual error, the steps in our guide to fixing what ChatGPT gets wrong about your company apply. When it is a fair criticism, the fix is the product or service, then a page that says clearly what changed and when.

Report framing per engine, against the same rivals

Report recommended share, the share of answers that place you at Top pick or Recommended, per engine and against the same named rivals you use for share of answer. A framing number without rivals cannot tell you whether the engines grew warmer to you or to the whole category. Our note on picking a rival set covers how to choose them.

Keep the engines apart. Gemini, ChatGPT, Copilot, Google AI Overviews and Google AI Mode draw on different sources and phrase judgements differently, so one blended framing score hides where the work is. A brand can be a Top pick in one engine and Qualified in another for the same prompt.

Show the reason field as a short table: the attributes each engine ties to you and to each rival, with counts. That table is often what a sales or product lead acts on, because it shows how the market is being told to see you.

Report it quarterly, on the same prompt set and the same coding definitions as the baseline. Changing the scale mid-year breaks the comparison as surely as changing the prompts. If the definitions must change, code the previous round again with the new ones so the trend line stays honest. Keep a short log beside the numbers: the date of each fix, each engine model change you know about, and each change to the coder. Most framing shifts that look mysterious in a chart are explained by one line in that log.

Which prompts deserve the most attention?

Commercial prompts, where a buyer asks for a provider to choose, carry the framing that moves money. Profound, which sells AI visibility tools, analysed 7.5 million US English ChatGPT conversations from 23 June 2025 to 22 June 2026.8 The share with commercial intent rose from 13.9% to 19.2% over that year.8 Tag your prompts by intent, and read the framing of the commercial ones first.

What should you do when an engine frames you as second choice?

Work out which reason the engine gives for the brand it puts first, then check whether your own pages make the same claim clearly enough to be read. Second place is usually a content gap, not a judgement against you. The winner is described in a way the engine can repeat, and you are not.

Start with the reason field for the rival above you. If the answer says they are best "for multi-country payroll", look at whether any page of yours states that you handle it, with which countries and since when. Engines repeat plain statements. A feature buried in a PDF or a sales deck does not get read.

Then look at third-party descriptions. In Semrush's survey of 622 US B2B professionals in March and April 2026, the 519 who use AI at work were asked what they do after an AI recommendation.9 Of those, 71% said they visit vendor websites and 38% check G2 reviews.9 The same people read the review sites the engines often draw on, so a weak profile hurts twice.

Finally, re-run the affected prompts after each fix and code them with the same scale. Our view: a fix is done when the framing has moved on the prompts it was aimed at, across more than one day, not when the page is published.

What a framing measure cannot tell you

A framing score tells you how the engines describe you on your prompt set. It does not tell you what every buyer saw, and it does not prove the description changed a decision. Buyers phrase questions differently, carry chat history and use engines in different countries. Building the prompt set from real buyer questions narrows that gap; it never closes it.

Accuracy problems are also broader than any one brand. The largest public review of assistant accuracy is now dated: the EBU and BBC study of news answers, collected in May and June 2025, found that 45% of AI answers had at least one significant issue.10 That covered news, not company facts, and engines have changed since. It is still a reminder that some wrong framing is noise in the engine, not a signal about you.

Treat the measure as a way to find and rank fixes. Pair it with share of answer for reach, the citation map for sources, and pipeline signals for money. Framing on its own is a diagnosis, and the treatment is still on your pages and profiles.

Start small. Take the prompts where you are already named, code one round by hand with the five fields, and look at the reason and caveat columns. The point of the first round is to find any description the team did not know the engines were repeating, and the page behind it.

Sources

  1. G2, 2026 AI Search Insight Report: 1,076 B2B decision makers, North America, EMEA and APAC, Mar 2026
  2. G2, 2026 AI Search Insight Report: 1,076 B2B decision makers, North America, EMEA and APAC, Mar 2026
  3. TrustRadius, Beyond the hype: 1,862 technology buyers, Jan 2026 (Jul 2026); 94% is of buyers who used AI
  4. Parse, Being cited is not enough: how to monitor sentiment in AI answers (2026)
  5. MaxAEO, AI visibility sample size and clustered runs (Jul 2026)
  6. Ahrefs, ChatGPT query fan-out study: 66.6M free, logged-out GPT-5.6 Luna responses, Aug–Sep 2026, all countries; factual subset 20,000 US prompts; citation ranks US, Mar–Sep 2026 (Oct 2026)
  7. Profound, where inaccurate AI claims come from (2026)
  8. Profound, commercial conversations in ChatGPT: 7.5M en-US conversations, 23 Jun 2025–22 Jun 2026, with industry splits (Aug 2026)
  9. Semrush, how AI shapes B2B buying: 622 US B2B professionals, Mar–Apr 2026; figures are of the 519 who use AI at work (Jul 2026)
  10. EBU and BBC, News Integrity in AI Assistants: answers collected May–Jun 2025 (Oct 2025)(dated)

Questions readers ask

  • Sentiment analysis alone is too coarse to measure how AI answers describe a brand, because many answers that name a company read as neutral. The useful signal is recommendation strength: whether the answer picks you, recommends you among a few, lists you, qualifies you or advises against you. Record the reason and any caveat as well, so each score points to something you can change.

Keep reading

Free. Enter your domain and Sigzen AI checks how five engines answer your buyers' questions, against up to three competitors.

No account needed. Two minutes to submit; we email the full five-engine report when the check has run.