Guides / Track brand mentions and sentiment
How to track brand mentions and sentiment in AI answers: a marketing manager's workflow
By Olly, founder of AEOscore, AEO consultant and full-stack developer, 20+ years in search14 min read
Track brand mentions and sentiment in AI answers by writing down 20 to 40 real buyer questions, asking them every week on ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews and AI Mode, and logging four things per answer: whether you were named, whether your site was cited, where you appeared, and how you were described on a fixed five-label scale.
TL;DR
- One check is a snapshot, not a measurement: no two of our 40 test runs returned the same brand list, so run each prompt more than once and report rates across weeks.
- Log the exact sentence the engine used about you. The wording is the evidence, and it is what you fix.
- Count “absent” as a sentiment result. It is the most common one, and the one most teams forget to report.
- Measurement is half the job. Each gap found should become a fix shipped and a re-ask four to twelve weeks later.
40
runs of one prompt: no two returned the same brand list
11
columns in the log, one row per answer
5
sentiment labels, including absent
4 to 12 wks
from a fix going live to answers moving
The log and rubric below are free to download, with no email gate: the mention and sentiment log template (Excel) or the plain CSV.
On this page
The metrics
What you are measuring in an AI answer
How can a digital marketing manager track brand mentions and sentiment inside LLM answers?
Build a fixed set of 20 to 40 real buyer questions, run them weekly across ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews and AI Mode, and log four things for every answer: whether your brand was mentioned, whether your site was cited, where you appeared in the answer, and how you were described on a fixed sentiment scale.
Four things, per answer. Each has its own glossary entry, and each moves on its own: a business can be mentioned without being cited, cited without being named, and named warmly on one engine while another hedges.
| Metric | The question it answers | How to count it |
|---|---|---|
| Mention rate | Does the answer name you? | Answers naming your brand divided by all answers to non-branded prompts. The headline number. |
| Citation rate | Does it link your site? | Answers citing a page on your domain divided by all answers. Usually about half the mention rate, and much lower on ChatGPT. |
| Answer position | Are you named first or fifth? | Where your brand sits in the order of brands named. First or second is a recommendation; sixth is a footnote. |
| Sentiment | How are you described? | The tone of the sentence around your name, scored on a fixed scale, with that sentence pasted into the log. |
What is the difference between a brand mention and a citation?
A mention is the AI naming your brand in the text of the answer, while a citation is the AI linking your website as a source, and a business can get one without the other.
The gap between the two is large. Similarweb’s US desktop data put outbound links in only 6.8% of ChatGPT answers as of May 2026 (Search Engine Journal), and in our own August 2026 scans 46% of ChatGPT answers that named a monitored brand carried no link to its domain, against 8% on Gemini (AEOscore KPI data). Count both, report both, and never let a citation dashboard stand in for a mention log. Named, cited or linked is the short version.
Why it is worth the effort: Forrester’s Buyers’ Journey Survey found 94% of business buyers used AI somewhere in their buying process in 2025 (Forrester, January 2026), and Ofcom’s April 2026 survey put 54% of UK adults on AI tools such as ChatGPT, Copilot or Gemini (Ofcom). Those answers replace clicks: Pew Research Center found US Google users clicked a traditional result on 8% of visits with an AI summary against 15% without, from March 2025 data (Pew Research Center), and our own AI Overviews click-through research collects the UK B2B picture. The answers those buyers read instead are the ones you are about to log.
The catch
Why you cannot just check once
Why do two people get different AI answers to the same question?
AI engines generate a fresh answer each time and vary by account history, location and model version, which is why a single check is not a measurement and every prompt should be run more than once.
SparkToro ran 12 identical prompts 2,961 times and found under a 1 in 100 chance that any two runs returned the same brand list (SparkToro, January 2026). We saw the same thing on a smaller scale: one UK B2B buyer prompt asked 20 times each on ChatGPT and Gemini on 20 August 2026, and no two of the 40 runs matched (our variance test). Re-running the same prompts on three separate days in August 2026 changed the brands named in 83% of prompt-and-engine combinations (our free-checker study).
The variation has three layers: between runs of the same prompt on the same engine, between engines (named on one is not named on all), and over time as sources and models change (answers change, ask again). A tracking process has to absorb all three, which is why the rest of this page is about repetition and consistency rather than cleverness.
“One check is a snapshot. A mention rate across three runs and four weeks is a measurement.”
Step 1
Build a prompt set of real buyer questions
How many prompts do you need to track brand mentions in AI answers?
Between 20 and 40 real buyer questions is enough for one business to produce a mention rate it can trust, split across category, comparison, problem-led and branded prompts, and frozen so that this month’s number is comparable with last month’s; agencies reporting to clients run larger sets of 30 to 100.
Write the questions the way a buyer types them, not the way your website describes you. Pull them from sales call notes, support tickets, Search Console queries and the questions your rivals’ pages answer. Split them across four types, and freeze the set once it is written: a prompt set that changes every month cannot produce a trend. The prompt set entry covers the freezing rule in more detail.
| Type | What it tests | Example prompts, written out |
|---|---|---|
| Category | The buyer knows the category and asks who is in it. Every answer is a list of brands, so this is where mention rate is won or lost. | “What are the best AI visibility monitoring tools for a UK B2B company?” “Which UK commercial cleaning companies work with multi-site retailers?” |
| Comparison | The buyer has a shortlist and wants it whittled. Engines lean on comparison pages and reviews here. | “Profound vs Peec AI for a small marketing team: which is better value?” “Xero or QuickBooks for a UK consultancy with three staff?” |
| Problem-led | The buyer describes the problem, not the product. These prompts find out whether the engines connect your name to the job you do. | “Why doesn't ChatGPT recommend my business?” “How do I stop AI Overviews taking clicks from my product pages?” |
| Branded | The buyer has heard of you and wants a verdict. These prompts measure sentiment and accuracy; keep them out of the mention rate. | “Is AEOscore any good?” “What do customers say about [your brand]?” |
Step 2
Decide your engines and cadence
How often should you check AI answers for brand mentions?
Weekly is the practical minimum for most B2B businesses because AI answers shift between runs and weeks, with daily checks only justified in fast-moving or highly competitive categories.
Track all six: ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews and Google AI Mode. They read different sources and disagree with each other, and the engine your buyers use is rarely the one you use. AI answers shift between runs and between weeks, so a monthly check cannot tell a trend from noise, and a daily check only pays for itself in fast-moving or heavily contested categories. ChatGPT only switched web search on for 34.5% of queries in February 2026, with monthly swings between 15% and 66.3% (Semrush), so even the same engine answers from different material week to week.
Ofcom counted 1.8 billion UK visits to ChatGPT in the first eight months of 2025, nearly five times the year before, and 53% of UK adults now say they often see AI summaries in search results (Ofcom, December 2025). Google AI Overviews and AI Mode belong on the list even if nobody in your office uses them.
Step 3
Run the manual version first
Can you track brand mentions in ChatGPT manually?
Yes, you can type your prompts into each engine by hand and record the results in a spreadsheet, but at six engines and three runs per prompt it stops being practical beyond a handful of prompts, because answers change between runs and need repeating; most teams sustain about 20 prompts only by cutting to one run on two engines.
Do it by hand for the first month even if you intend to buy a tool. You will learn what the answers in your category look like, which sources keep appearing, and what a “mention” means in practice. The procedure:
- Open a fresh, logged-out or temporary chat on each engine so memory and custom instructions cannot personalise the answer.
- If you sell to one country, use a VPN or a browser located there, because the engines answer differently by region and AI Overviews trigger on different queries in different countries.
- Paste the prompt exactly as written in your set. Do not rephrase it between runs.
- Record the answer before you react to it: who was named, in what order, which sources were linked, and the sentence that mentions you.
- Repeat the prompt in a new chat. Three runs per prompt per engine is the practical minimum.
- Move to the next engine and do the same.
The limit is arithmetic. 20 prompts, 6 engines and 3 runs each is 360 answers a week to read and score. At 2 minutes each, that is about 12 hours a week, every week, and the regional and logged-out caveats make it slower than it sounds. Most people manage it once at full rigour, then cut to one run on two engines, or do it monthly, which is a snapshot, not a trend.
Step 4
Log it consistently
What should an AI brand mention log contain?
Eleven columns per answer: date, engine, prompt text, mentioned yes/no, named first yes/no, cited yes/no, sentiment label, the exact sentence the AI used, competitors mentioned, sources the AI cited, and notes.
One row per answer, every time, with these columns. The template below has the same header row, the rubric from step 5, and a scored example of each label.
| Column | What goes in it |
|---|---|
| date | The day you ran it, as 2026-10-12 |
| engine | ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews or Google AI Mode |
| prompt text | The prompt exactly as asked |
| mentioned yes/no | Was your brand named anywhere in the answer text? |
| named first yes/no | Was yours the first brand named? |
| cited yes/no | Was a page on your domain linked as a source? |
| sentiment label | positive, neutral, negative, mixed or absent |
| exact sentence the AI used | Paste the sentence or clause that mentions you, verbatim |
| competitors mentioned | Every other brand named, in the order named |
| sources the AI cited | The domains linked or listed as sources |
| notes | Anything odd: a wrong fact, an old price, a search that did not run |
Free template, no email gate
The mention and sentiment log
The eleven-column log with the five-label rubric on a second tab and one made-up example row per label, so a colleague can take over the scoring without a briefing.
- Log tab: the eleven columns, one row per answer
- Dropdowns for engine, yes/no and the sentiment label
- Rubric and rules tab: the five labels with an example each
- Six made-up example rows to copy, then delete
Free to use and share. Works in Excel, Numbers and Google Sheets.
“Paste the sentence, not your summary of it. The engine’s wording tells you which source to fix.”
Step 5
Score sentiment with a published rubric
What is AI sentiment and how is it scored?
AI sentiment is the tone an AI answer uses when it describes your brand, scored against a fixed scale of positive, neutral, negative, mixed or absent, with the exact sentence the engine used recorded as evidence.
Five labels, one of them “absent”. Absent matters as much as negative: an answer that recommends three rivals and not you is a sentiment result about you, and it is the most common one. Paste the sentence the engine used into the log verbatim. The wording is the evidence, and it usually points at the source the engine was repeating.
| Label | When to use it | Example answer (invented) |
|---|---|---|
| Positive | The engine recommends you or attaches a favourable attribute to your name. | “Brightwater Payroll is a strong choice for UK firms under 50 staff: HMRC-recognised, with same-day support.” |
| Neutral | You are listed or described factually with no judgement either way. | “Other options include Brightwater Payroll, Ledgerly and PayPilot.” |
| Negative | The engine warns against you, attaches a caveat, or repeats a complaint. | “Brightwater Payroll has no pension auto-enrolment module, so check it covers your scheme first.” |
| Mixed | A recommendation and a caveat in the same answer. Log it as mixed, not as an average. | “Brightwater Payroll is well reviewed for support, although its reporting is thinner than Ledgerly’s.” |
| Absent | The answer names rivals, or nobody, and not you. Count it; it is the result most teams forget. | “The leading options are Ledgerly, PayPilot and Crestline Payroll.” |
“Absent is the most common sentiment result and the one most teams never count. You cannot fix a gap you have not logged.”
Step 6
Handle variance
How do you handle AI answers that change between runs?
Run every prompt at least three times per engine in fresh sessions and score the prompt on the share of runs that named you, so a brand named in two runs out of three is logged as a 67% mention rate rather than a yes or a no.
We use 3 runs per prompt per engine as the working minimum. Score the prompt on the share of runs that named you, never on the last run you happened to see. A prompt that names you in two runs out of three is a 67% mention rate for that prompt and engine this week. Report it as that, with the three sentences logged, and watch whether it holds next week.
The number that matters is the rate across runs and weeks, not any single answer. SparkToro’s finding cuts both ways: single runs are near-random, but aggregate frequency across many runs is stable enough to track (SparkToro). Our own scans weight mention rates with confidence intervals for the same reason; the methodology page explains how.
Step 7
Turn the log into a report your boss reads
Which metrics should go in an AI visibility report?
Report four numbers, mention rate, citation rate, named-first rate and the sentiment split, each as a trend across weeks rather than a single snapshot, and add competitor share of voice once you have a month of data.
Four numbers, one chart, one list. Anything longer gets skimmed.
Four numbers
- Mention rate, non-branded prompts only, per engine and blended
- Citation rate, per engine
- Named-first rate, the share of mentions where you were the first brand
- Sentiment split: positive, neutral, negative and mixed as shares of the answers that name you, with absent reported alongside as the share of all answers that do not
One chart
One line per engine, mention rate by week, with the dates your fixes went live marked on it.
One list
The prompts where you were absent or negative on most engines, each with the sources those answers cited. That list is next month’s work.
Add competitor share of voice once you have a month of data; the formula and a worked example are in the AI share of voice guide. If you report to clients rather than a boss, the AEO KPIs guide has the nine KPIs with formulas and a monthly report template.
Step 8
Act on what you find
Does tracking sentiment change anything by itself?
Tracking only tells you where you are missing or misdescribed, so the value comes from the fixes that follow: new or rewritten pages, corrected facts on third-party sources and better citable evidence, which typically take four to twelve weeks to show in answers.
Most tracking pages stop at the dashboard. The log is only useful if every absent or negative row becomes a fix, and every fix gets re-asked. The loop is: gap found, fix shipped, re-ask weekly, judge at week eight, decide at week twelve.
| What the log shows | The fix | Re-ask |
|---|---|---|
| Absent on a category prompt, rivals named | Read the sources those answers cite. Get onto the roundups and directories doing the recommending, and publish the page that answers the question in its first sentence. | Weekly. First movement on live-search engines in days to four weeks; judge at week eight. |
| Negative or mixed on a branded prompt | Find the quoted source. Fix what is fixable there (an old review, a stale listing, a wrong price), then publish the corrected fact plainly on your own site. | Weekly. Sentiment moves when its evidence moves, usually within four to twelve weeks. |
| Mentioned but never cited | Check the crawlers can read the page, put the answer in plain HTML near the top, and show a real updated date. | Weekly. A citation can appear before the brand is named in the answer text, so log both. |
| Named fifth behind rivals | Publish the fair comparison page naming those rivals, with prices, and get your name into the same roundups that put them first. | Weekly. Position moves slowly; give it the full twelve weeks. |
How long it takes, from our own log
We shipped 8 fixes on our own site from 17 to 24 August 2026 and followed each one through weekly scans to 21 September 2026. 63% of the fixes got us named within five weeks, the median wait for a first mention was 5 days among fixes that got one, and 25% turned into repeat mentions. ChatGPT and Perplexity moved first; Google AI Overviews took four weeks. The full table and the four-phase timeline are in how long does AEO take (fix-to-mention log).
Which fixes, in which order, is covered in what gets fixed after an AEO audit, and the difference between tools that measure and providers that fix is the whole of monitoring vs fixing. Schema markup is housekeeping here, not a lever: it helps machines parse the page you already have and does not raise mentions on its own.
The threshold
When to stop doing it manually
When should you stop tracking AI mentions manually?
Stop when prompts multiplied by engines, runs and weeks adds up to more than a working day a month of reading and scoring, which at six engines and three runs happens well before 20 prompts, or when you need week-on-week trend data rather than one-off snapshots.
Multiply prompts by engines by runs by four weeks, at about two minutes an answer. Once reading and scoring the answers takes more than a working day a month, the log is costing you more than a tool would. At six engines and three runs, a working day a month covers about three prompts; at one run on two engines, about 30. Done properly, the point arrives long before the 20 to 40 prompts the method needs; done loosely, you are measuring noise. Below the threshold, the spreadsheet is the right tool.
AEOscore is a UK AI visibility monitoring platform that runs this workflow for you and then fixes what it finds: your buyer questions asked every week on all six engines (every day on Monitor + Strategist), every answer scored for mention rate, citation rate, answer position, competitors named and sentiment (positive, neutral or negative, with the sentence kept as evidence), the sources each answer cites recorded, share of voice per prompt, a written fix for each gap, and a 7 and 28 day re-check on every fix you ship. On Monitor + Strategist an in-house strategist and developer makes the fixes, so each one goes live the same month. Pricing in one line: £395 one-off AI Visibility Audit, credited against your first month, £99 a month Monitor, £1,495 a month Monitor + Strategist. Prices exclude VAT; details on the pricing page. The free AEO score gives you a week-zero baseline across all six engines before you spend anything, and the comparison table now shows how each tool handles mention and sentiment tracking.
FAQ
Tracking AI mentions and sentiment: common questions
- Do you need a tool to track AI sentiment?
- No tool is required to start, but a tool becomes worth paying for once manual checking takes more than a working day a month or you need week-on-week trend data rather than one-off snapshots.
- Should branded prompts count towards the mention rate?
- No: a prompt that contains your own name will almost always get you mentioned, so keep branded prompts in the set for sentiment and accuracy but report mention rate on the non-branded prompts only.
- Is a neutral mention a good result?
- Usually yes for a category prompt, where being listed at all is the win, and not enough for a branded prompt, where the buyer asked for a verdict and the engine declined to give one.
- Can you use the free ChatGPT plan to run the checks?
- Yes, as long as you use a temporary chat or log out so memory and custom instructions cannot personalise the answers, and you note which model you were on.
- Why does Google AI Overviews not appear for some of my prompts?
- AI Overviews only show for some queries, so log those prompts as no overview shown rather than absent, and track whether an overview starts appearing as the category heats up.
- What counts as a mention if the engine gets your name slightly wrong?
- Count it as a mention and note the variant in the notes column, because engines often reuse a misspelling from a third-party source and the fix is to correct that source.
- How does AEOscore score sentiment?
- AEOscore reads every scanned answer that names you and labels it positive, neutral or negative, stores the sentence it based the label on as evidence, counts absence through the mention rate, and groups the negative and neutral answers by prompt with a written diagnosis and fix for each.
Sources
Numbered sources
- Forrester (John Buten), B2B Buyers Make Zero-Click Buying Number One. Accessed 12 October 2026. 94% of business buyers report using AI in their buying process (Forrester Buyers' Journey Survey, 2025). Published 22 January 2026.
- Ofcom, Adults' Media Use and Attitudes 2026. Accessed 12 October 2026. 54% of UK adults use AI tools such as ChatGPT, Copilot or Gemini. Published 2 April 2026.
- Ofcom, From apps to AI search: how the UK goes online in 2025 (Online Nation 2025). Accessed 12 October 2026. ChatGPT had 1.8 billion UK visits in the first eight months of 2025, up from 368 million in the same period of 2024; 53% of UK adults say they often see AI summaries in search results. Published 10 December 2025.
- SparkToro (Rand Fishkin), AIs are highly inconsistent when recommending brands or products. Accessed 12 October 2026. 2,961 runs of 12 identical prompts: under a 1 in 100 chance that two runs return the same brand list, while aggregate frequency across many runs is far more stable. Published 28 January 2026.
- AEOscore, Run-to-run variance test. Accessed 20 August 2026. One B2B buyer prompt asked 20 times each on ChatGPT and Gemini in fresh sessions on 20 August 2026: no two of the 40 runs returned the same brand list.
- AEOscore, Free AI visibility checkers for UK businesses: how much answers change between days. Accessed 20 August 2026. The same buyer prompts re-run on three separate days in August 2026: the brands named changed in 83% of prompt-and-engine combinations.
- Pew Research Center, Google users are less likely to click on links when an AI summary appears in the results. Accessed 12 October 2026. US users clicked a traditional result on 8% of visits with an AI summary against 15% without. 900 adults, 68,879 searches, March 2025. Published 22 July 2025.
- AEOscore, AI Overviews and click-through rates for small UK B2B. Accessed 12 October 2026. Seer, Pew and Ahrefs click-loss figures attributed and dated, plus our own 304-query UK B2B study of which query types trigger an AI Overview (17 August 2026).
- Search Engine Journal, reporting Similarweb's 2026 Generative AI Landscape report, ChatGPT Links Out Most On Travel Queries, Data Shows. Accessed 12 October 2026. 6.8% of ChatGPT answers included outbound links as of May 2026. US desktop usage; Similarweb describes its figures as estimates. Published 27 July 2026.
- AEOscore, AEO KPIs that replace keyword rankings: first-party figures. Accessed 12 October 2026. August 2026 scans across the sites AEOscore monitors: 22% of answers naming a monitored brand carried no link to its domain, 46% on ChatGPT and 8% on Gemini.
- AEOscore, Fix-to-mention log for aeoscore.co.uk. Accessed 12 October 2026. Eight fixes shipped 17 to 24 August 2026, tracked to 21 September 2026 across five engines: 63% of fixes got us named within five weeks, median 5 days to a first mention among fixes that got one.
- Semrush, ChatGPT traffic analysis: insights from 17 months of clickstream data. Accessed 12 October 2026. ChatGPT enabled web search on 34.5% of queries as of February 2026, with month-to-month swings between 15% and 66.3%. US clickstream panel. Published 7 April 2026.
Get your week-zero baseline
The free AEO score runs your buyer questions across all six engines, human-checked, within 1 to 2 working days. Domain and email, no card.
- ChatGPT
- Claude
- Gemini
- Perplexity
- Google AI Overviews
- Google AI Mode