Skip to content

Integration

AI brand sentiment API: how to measure tone in AI answers

Eight steps to score how AI answers describe your brand, with the field each API returns and a bill from $2.94 per 1,000 collected and scored answers.

Five isometric cubes in a row with the center cube highlighted in blue

SEO consultant, and more recently GEO, at MentionBox

To measure brand sentiment in AI answers with an API, you run a fixed prompt list on the engines, store every answer, and attach a polarity to each mention of your brand, either by buying an API that returns the label or by classifying the stored text yourself. No engine returns the figure: ChatGPT, Gemini, Claude and Perplexity hand back text, searches and cited URLs, and every sentiment score in a GEO report is computed on top of that. Published rates for collecting and scoring 1,000 answers run from $2.94 on Rankability to $81.94 on AthenaHQ, while classifying answers you already own costs under ten cents per 1,000 in model tokens. API for GEO is our editorial pick after hands-on testing by Mathias C., at $13 per 1,000 answers with the score, the label, the themes and the per-brand perception in the same response.

This guide is for a brand team that has a visibility number and now needs to say how the engines talk about the brand, and for the developer who will write the job. Steps 1 and 2 settle what you measure and at what grain, steps 3 and 4 pick the API and show the requests that return the field, steps 5 to 8 cover rolling your own classifier, the aggregation arithmetic, the sample size and the bill. MentionBox, which publishes this site, has a commercial link with API for GEO.

What a brand sentiment API actually returns

Three layers come back, and the vendor you pick decides how many of them you can see:

  • The answer text. The sentence that describes your brand is the evidence. Without it, a label is an assertion nobody in the room can check.
  • The label or the score. Positive, neutral, negative, mixed, or a number on a published scale. This is the layer that differs most between vendors, in shape as well as in value.
  • The reason. An aspect, a theme or a claim that says what the tone was about: price, support, reliability. This is the layer that turns a chart into work.

Sentiment is the loosest defined of the GEO metrics, and our explainer on GEO metrics covers the denominators that differ on the share of voice side. A share is a division you can audit. A sentiment label is a judgement on a sentence, and judgements vary between vendors, between models and between two humans reading the same line.

One distinction saves a quarter of confusion. Social listening measures what people wrote about you on platforms you can crawl. A brand sentiment API measures what a model wrote about you when someone asked it a buying question, in text that was never published and that nobody else can retrieve. The two numbers answer different questions and will not agree.

1. Write prompts that can carry sentiment

A prompt list built for visibility measures appearance. Sentiment needs prompts that make the engine commit to an opinion.

Four families earn their place in the list:

  • Recommendation prompts. "Best accounting software for a five-person agency", "which CRM should a B2B startup pick". Brands get named and ranked, and the reasons come with them.
  • Direct brand prompts. "Is [your brand] worth the price", "what are the downsides of [your brand]". These produce the most negative labels in any list, which is the point.
  • Comparison prompts. "[Your brand] versus [competitor]", "alternatives to [your brand]". The framing you get here is the framing a buyer reads at the shortlist stage.
  • Avoidance prompts. "Which tools in this category should I avoid", "why do teams leave [competitor]". Unflattering claims surface here first, usually months before they reach a recommendation answer.

Tag every prompt as branded or unbranded before the first run. The distinction is built into the APIs: Conductor's sentiment endpoint carries a prompt_type field whose values are branded and unbranded, and Promptwatch types each prompt as ORGANIC, BRAND_SPECIFIC or COMPETITOR_COMPARISON. A sentiment average that mixes the two is two populations in one line: branded prompts almost always score warmer, because the engine is answering a question about you rather than choosing between you and four others.

2. Choose the unit you score

Five units are in production across the APIs we track, and they are not interchangeable. The unit decides what you can report and what you can fix.

Unit scoredWhat one record isPublished example
The answerone polarity for the whole responseScrunch brand_sentiment: positive, mixed, negative, none
The mentionone polarity per time the brand is namedAhrefs Brand Radar brand_sentiment[]: brand, sentiment, aspect, snippet
The claimone polarity per statement made about the brandProfound sentiment_claims: asset, theme, claim, sentiment
The attributeone polarity per praised or criticised propertyAthenaHQ attribute observations with a score from -1 to 1
The brand, over a windowone aggregate per brand per periodOtterlyAI sentiment: positive, neutral, negative, nss

Field names and allowed values from each vendor's API reference, read on November 6, 2026, with prices and plans from the same vendors checked on October 23, 2026.

The practical rule is that you can always aggregate downwards and never upwards. Per-mention and per-claim records let you compute a response-level polarity, a per-aspect split and a brand aggregate from the same dataset. A brand aggregate cannot be broken back apart, so the month a client asks whether the negativity is about price or about onboarding, the answer has to be recollected.

The grain also decides whose problem a bad number is. Conductor's nine sentiment_category values, misc, quality, price, service, user-friendliness, experience, popularity, health and ethics, map almost one for one onto internal owners. Ahrefs' eleven aspect values do the same job: features_specs, performance_effectiveness, price_value, reputation_popularity, quality_reliability, design_comfort, usability_compatibility, availability_delivery, safety_health, service_support and general.

3. Pick an API that publishes the sentiment field

Twenty of the 55 APIs we rate return a sentiment on brand mentions. The twelve below publish enough about the field to integrate against, and the first row is what we would build on. Our ranking of GEO metrics APIs scores the whole group, and the ranking for enterprise brand teams reweights the same criteria the way a brand team pays for them.

APIEngines, of the 10 we trackHow sentiment comes backPrice per 1,000 answersEntry ticket that carries the keyResale
API for GEO10sentiment object: score 0 to 100 and its label, per brand, plus theme clusters and per-brand perception$13, down to $7 at volumepay as you go, no minimum publishedallowed
LLM Pulse10sentiment per answer and per prompt, with a negative-sentiment webhook€30.67€299 a month, Scaleallowed
Promptwatch8per response and per brand: sentiment POSITIVE, NEGATIVE, NEUTRAL, plus sentimentScore and toneOfVoice$15.83$95 a month, Essentialallowed
Rankability10brand sentiment on each AI row, with a mention_status per brand$2.94$97 a month, Starternot allowed
AthenaHQ10attribute observations: praise or criticism with a score from -1 to 1$81.94, plus an API add-on priced on request$295 a month, Starterallowed
Scrunch9per response: positive, mixed, negative, none; brand_sentiment_score 0 to 100 on the Query APInot computable$500 a month, Agency Corenot disclosed
OtterlyAI7per brand per window: positive, neutral, negative counts and an nss$15.75$189 a month, Standardnot disclosed
Peec AI9one 0 to 100 metric with its aggregation formula publishednot publishedPro and above, price on requestnot allowed
Profound9positive_sentiment and negative_sentiment percentages, grouped by theme or claimnot publishedEnterprise plan on quotenot allowed
Ahrefs Brand Radar8per mention: brand, label among four, aspect among eleven, and the sentence$20$129 a month, Lite, plus check packsnot allowed
SE Visible4positive, neutral or negative per prompt result$3.30$99 a month, Basicnot allowed
Conductor8one row per sentiment mention: value, score, category, snippet and source URLnot publishedConductor Intelligence subscriptionnot disclosed

Prices, engine counts and plan floors from each vendor's pricing page and API reference, checked on October 23, 2026. Rates per 1,000 answers are our arithmetic on the published allowance of the cheapest plan that carries an API key. For API for GEO, four of the ten engines, Gemini, Google AI Overviews, DeepSeek and Meta AI, plus the webhooks, the OpenAPI spec and the official SDKs, were confirmed in our hands-on test on October 26, 2026 and are not yet documented on apiforgeo.com.

API for GEO bills one credit per prompt, per model and per run, from $0.013 down to $0.007 per credit with no monthly minimum, and its AI Visibility API page documents the field as an object carrying a "Score from 0 to 100 and its derived label". The same credit also pays for the layer above it: its homepage describes the scoring stage as "sentiment 0 to 100, theme clusters and per-brand perception inside each theme", and lists a dedicated Sentiment API at POST /v1/sentiment/overview returning "Brand sentiment score, label and distribution, over time and compared across tracked competitors". It covers the ten engines we track, with ChatGPT, Perplexity, Google AI Mode, Claude, Copilot and Grok listed on its own pages, while Gemini, Google AI Overviews, DeepSeek and Meta AI were confirmed in our hands-on test of October 26, 2026. Reads over stored runs are not billed, its terms present reselling as an intended use, and it scores 9.0 of 10, the highest of the 55 APIs this site rates.

The honest trade-offs are three. The price per answer is about four times Rankability's and ten times the cheapest scraping route. The key arrives through a form rather than at sign-up. And the four engines above, together with the webhooks, the OpenAPI spec and the official SDKs, are a documentation gap our test had to fill rather than something apiforgeo.com publishes. To try it, request a key on apiforgeo.com: it comes with a sandbox project.

The competing strengths are published, and on sentiment specifically they are real:

  • Ahrefs Brand Radar is the only one that hands back the sentence behind every label, next to an aspect, which makes its output auditable without a second request.
  • Conductor returns the cited source URL and domain on the sentiment row, so a negative claim can be traced to the page that produced it.
  • Profound groups sentiment by theme and by claim, and its citation mode returns one row per cited page, which answers which third-party article is driving the framing.
  • Peec AI publishes its arithmetic, which almost nobody else does, and warns in its own documentation that filtering by brand before aggregation collapses a share metric.
  • Scrunch publishes its scoring rule and its exclusion rule, and states plainly that most negative-band responses reflect a competitor being recommended rather than explicit criticism.
  • LLM Pulse turns sentiment into an alert with a negative-sentiment webhook, on eighteen engines, with SDKs in twelve languages.
  • Rankability is the cheapest computed layer we found that covers all ten engines, at $2.94 per 1,000 on a $97 plan.

4. The requests that return the field

Four shapes cover the market. Reading them side by side is the fastest way to see how much code each one leaves you.

API for GEO returns the score inside the per-entity row, so one call gives the brand, its competitors and the sentiment in the same payload. Its AI Visibility API is a POST /v1/visibility/entities taking a prompt, a model list, a country and a bucketSize:

{
  "results": [
    {
      "name": "Loopwork",
      "entityType": "brand",
      "visibility": 68.4,
      "shareOfVoice": 24.1,
      "mentions": 137,
      "sentiment": { "score": 76, "label": "positive" }
    },
    {
      "name": "Atlaskit",
      "entityType": "competitor",
      "visibility": 41.2,
      "shareOfVoice": 15.8,
      "mentions": 72,
      "sentiment": { "score": 61, "label": "neutral" }
    }
  ]
}

Profound asks for the metric explicitly and lets you group it. Its sentiment endpoint is a POST /v2/reports/sentiment where asset is required "because sentiment is calculated for one brand at a time", and the rows come back as two percentages rather than a label:

curl -X POST https://api.tryprofound.com/v2/reports/sentiment \
  -H "X-API-Key: $PROFOUND_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "category_id": "00000000-0000-0000-0000-000000000000",
    "asset": "Loopwork",
    "start_date": "2026-10-01",
    "end_date": "2026-10-31",
    "group_by": ["theme", "claim"],
    "metrics": ["positive_sentiment", "negative_sentiment", "occurrence"],
    "limit": 50
  }'

A row answers with positive_sentiment and negative_sentiment as percentages of one another, an occurrence count when you ask for it, and the theme and claim that produced them. Setting source to citation switches the same endpoint to one row per cited page, with a citation_share, at the cost of refusing any group_by. The duel API for GEO vs Profound sets the two payloads side by side.

Conductor is the one to copy if you want rows you can load straight into a warehouse. Its ai_search_sentiments endpoint is an async POST, returns 29 positional columns with a schema block, and states that "each row represents a single sentiment mention for a brand or product within an AI response, including optional snippet and its source":

curl -X POST https://api.conductor.com/data-api/v1/async/ai_search_sentiments \
  -H "Authorization: Bearer cond_$CONDUCTOR_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "account_id": 123,
    "start_date": "2026-10-04",
    "end_date": "2026-10-31",
    "collection_frequency": "WEEKLY",
    "sentiment_values": ["negative"],
    "sentiment_categories": ["price", "service"]
  }'

Two details in that call will bite on the first attempt. A weekly window has to start on a Sunday and end on a Saturday, and a monthly one on the first and last day of the month. And the request is submitted, polled and paginated against the same path, with the executionId valid for seven days and parallel pagination refused.

Ahrefs puts sentiment in the answer listing rather than in a metric endpoint. Its Get AI responses endpoint is a GET on /v3/brand-radar/ai-responses with a required select and a required data_source, one date at a time rather than a range:

curl -G https://api.ahrefs.com/v3/brand-radar/ai-responses \
  -H "Authorization: Bearer $AHREFS_API_KEY" \
  --data-urlencode "select=question,brand_sentiment,links,data_source" \
  --data-urlencode "data_source=chatgpt,perplexity" \
  --data-urlencode "date=2026-10-31" \
  --data-urlencode "country=us" \
  --data-urlencode "limit=1000"

Its own cost note is worth reading before you widen the select: requests returning only custom prompt data are free, requests that include Ahrefs prompt data follow standard unit pricing, and three fields carry a 10-unit tag each, response, links and volume.

5. Classify the answers yourself, and what it costs

If you already collect answers through a scraping API, the sentiment layer is a second pass over text you own. The arithmetic makes it tempting: at the published OpenAI rates of $0.05 per million input tokens and $0.40 per million output tokens for gpt-5-nano, classifying one answer of about 1,000 tokens with a 300-token rubric and a 60-token answer costs about $0.00009, so about $0.09 per 1,000 answers. The batch table halves both rates, which takes it to roughly $0.05.

Route for the scoring passPublished rateCost to score 1,000 stored answers
gpt-5-nano, synchronous$0.05 in, $0.40 out per 1M tokensabout $0.09
gpt-5-nano, batch$0.025 in, $0.20 out per 1M tokensabout $0.05
gpt-4o-mini, synchronous$0.15 in, $0.60 out per 1M tokensabout $0.23
Claude Haiku 5.5, prompts up to 100,000 tokens$0.10 in, $0.50 out per MTokabout $0.16
Claude Haiku 5.5, Batch API$0.05 in, $0.25 out per MTokabout $0.08

Our arithmetic on the rates published by OpenAI and Anthropic, read on November 6, 2026, under one stated assumption: 1,300 input tokens and 60 output tokens per answer. Anthropic notes that its models from Claude 4.7 onwards use a tokenizer producing "approximately 30% more tokens for the same text", so budget the Haiku lines accordingly.

The rubric decides whether that saving is real. Four choices separate a label you can defend from a number that moves whenever someone edits the prompt:

  1. Fix the classes, and write the boundary cases. Three classes with a separate mixed label is the shape most vendors ship, and it is also the shape that disagrees most. Decide in writing where "cheaper than its competitors" lands, and whether a brand listed without comment is neutral or absent.
  2. Score the span, not the answer. Pass the sentence or the paragraph around the mention, with the full answer available as context, and store the span with the label. Ahrefs ships a snippet field and Conductor a sentiment_snippet for the same reason.
  3. Force the output shape. A JSON schema with an enum removes a whole class of parsing bugs and makes re-runs comparable.
  4. Keep a human sample. Fifty hand-labelled mentions per quarter, compared against the classifier, is the only way to know whether the series is measuring the category or your own prompt edits.

A minimal output contract looks like this:

{
  "type": "object",
  "additionalProperties": false,
  "required": ["brand", "polarity", "aspect", "span"],
  "properties": {
    "brand": { "type": "string" },
    "polarity": { "enum": ["positive", "neutral", "negative", "mixed"] },
    "aspect": { "enum": ["price", "quality", "support", "features", "availability", "reputation", "other"] },
    "span": { "type": "string" }
  }
}

Keyword lists are the cheaper route and the one to avoid as a final answer. "Not reliable" and "reliable" share a keyword, and a published comparison found the lexicon tools at the bottom: against a human majority vote on 100 tweets, a study published on arXiv on October 7, 2026 measured a Cohen's kappa of 0.37 for TextBlob and 0.40 for VADER on three-class sentiment, against 0.73 for a transformer model fine-tuned on tweets and 0.48 to 0.63 for the three open LLMs tested.

That same study sets expectations for the whole exercise, and it is the figure missing from every page we read on this subject. Six human raters labelling the same 100 tweets reached a Fleiss' kappa of 0.39 on three classes, which its authors describe as "only fair agreement, highlighting the subjectivity of sentiment analysis". Collapsing the task to negative against non-negative moved the humans to 0.45 and the fine-tuned model to 0.88 against the majority vote. The texts were tweets rather than AI answers, so the numbers transfer as an order of magnitude rather than a benchmark. The lesson transfers exactly: a two-class question is measurably more reliable than a three-class one, and "what share of mentions are negative" is a better headline than a positivity score.

6. Turn labels into one number you can defend

Take a month of 800 collected answers, 50 prompts on four engines at four runs each. Your brand is named in 304 of them: 150 positive, 118 in the middle band, 36 negative. Four published shapes turn those same 304 labels into four different headline figures.

Published shapeArithmetic on the same 304 labelsResult
Net sentiment score, positives minus negatives over mentions(150 − 36) ÷ 304 × 100+37.5
Scrunch rule: positive 100, mixed 50, negative 0, averaged(150×100 + 118×50) ÷ 30468.8
Peec AI formula on a −1 to +1 mean((0.375) ÷ 2 + 0.5) × 10068.8
Share of mentions that are negative36 ÷ 30411.8%

Our arithmetic, on the scoring rules published in the Scrunch metrics reference and the Peec AI sentiment page, read on November 6, 2026. Counts are illustrative and arithmetically consistent.

Two of those four are the same number arrived at by different routes, which is reassuring, and the other two describe the same month as a cheerful 68.8 or a specific 11.8% of mentions to go and fix. Scrunch publishes the warning in its own documentation: a score of 85 "does not mean '85% positive'", and the distribution only shows up if you break the figure down by band.

Three properties of this metric are worth writing into the report itself:

  • It is conditional on being mentioned. Scrunch averages over "responses that mention your brand", and Peec AI describes its score as the tone of the AI responses that mention your brand, so a brand named twice in glowing terms outscores a brand named 200 times with one recurring criticism. Print the mention count next to the score, every time.
  • Exclusions move it. Scrunch excludes its none band from the average rather than counting it as a middle value, which lifts the score whenever a month produces many flat mentions. A formula that scores the middle band at 50 and one that drops it are not comparable.
  • Per engine, never averaged. Engines are named in different proportions and answer in different registers, so a single cross-engine figure hides the surface where the framing is worst. Our own position, for what it is worth, is to report one line per engine and never a mean of means, for the same weighting reason that applies to visibility.

7. Set the sample size and the alert threshold

A sentiment score moves on its own. cloro's published study on sample size, dated August 14, 2026, measured a mean overlap of 0.40 between the cited domains of two runs of the same prompt on the same model minutes apart, with per-prompt overlap from 0.06 to 1.00. It puts about 300 answers per period at roughly ±5 percentage points, and about 294 answers on each side to support a claimed move from 20% to 30%.

Sentiment sits inside a smaller sample than visibility, because it is computed only where the brand appears. In the 800-answer month above, the score rests on 304 labels, and a per-engine split leaves about 76 labels per engine, which is a band of roughly ±10 points. Three rules follow:

  • Run each prompt at least four times a month, and compare quarters rather than weeks. Report the per-engine figure only where the mention count supports it.
  • Alert on the negative share and on new claims, not on the composite score. A score drifting two points is noise; a claim about pricing appearing in nine answers where it appeared in one is a fact with a date.
  • Log the collection route with every answer, and the engine, the country and the language. A month that mixes an interface capture with a model API call measures two populations, as our comparison of the OpenAI web search tool against ChatGPT answers sets out.

8. Put a price on the job

The volume is prompts × engines × runs × countries. For 50 prompts, four engines, four runs a month, one country, that is 800 answers.

RouteMonthly bill for 800 collected and scored answers
API for GEO$10.40 at $0.013 a credit, score, label, themes and perception included
Rankability$97, the plan floor; its Starter allowance covers far more than 800
SE Visible$99, the plan floor, on four engines
Promptwatch$95, the Essential floor, 6,000 answers included
Ahrefs Brand Radar$129 Lite, plus $50 for a 2,500-check pack to cover the 650 checks beyond the 150 included
OtterlyAI$189, Standard, the first plan with an API
LLM Pulse€299, Scale with weekly tracking
AthenaHQover $295, Starter, with the API add-on priced on request
Scrunch$500, Agency Core, no published answer volume
Profound, Peec AI, Conductoron quote
Collect on DataForSEO and classify yourselfabout $0.96 of scraping on the standard queue after a $50 minimum payment, plus about $0.07 of gpt-5-nano tokens, plus your own code

Our arithmetic on the rates and allowances each vendor published, checked on October 23, 2026, under three stated assumptions: a 30-day month, every prompt running on every included engine, and one country per prompt. Subscriptions collect more than 800 answers at their own frequency and bill their floor either way.

The last line is the one that decides most builds. Collection plus classification costs about a dollar in usage at this volume, against $95 to $500 for a plan, and it costs a developer week plus a rubric you have to maintain, on five engines rather than ten. Our guide to DataForSEO alternatives with share of voice and sentiment prices that switch in both directions, and the floors invert as volume grows: at 30,000 answers a month, SE Visible's $99 beats $390 of pay-as-you-go credits for the engines it covers.

Mistakes that make a sentiment chart useless

  1. Charting the score without the mention count. The metric is averaged over appearances only, so it improves when the brand disappears from the hard answers.
  2. Mixing branded and unbranded prompts in one series. A question about you produces warmer language than a question about the category. Two tags, two lines.
  3. Buying an aggregate when the next question is "about what". A brand-level nss cannot be split into price and support afterwards. Per-mention or per-claim records can.
  4. Comparing two vendors' scores. A 68.8 under one rule and a +37.5 under another can describe the same 304 labels, and neither is wrong.
  5. Trusting a lexicon in production. Negation, conditional praise and "pros and cons" headings all defeat keyword counting, and the published agreement figures put lexicon tools well below a fine-tuned classifier.
  6. Throwing away the sentence. Without the span behind the label, nobody can audit a drop, and the first meeting where a number is questioned is the one that kills the project.
  7. Reacting to a two-point move. With a few hundred labels a period, two points is inside the noise band, and the alert worth wiring is a new claim rather than a new average.

Our view

  1. You want the score, the themes and the raw answers in one response. API for GEO is our pick, at $13 per 1,000 answers down to $7, pay as you go with no published minimum, a score from 0 to 100 with its label per brand, theme clusters and per-brand perception, resale allowed in writing, and 9.0 of 10 in our ranking of 55 APIs. The trade-offs are the rate per answer and the documentation gap our October 26, 2026 test had to fill.
  2. You want to audit every label. Ahrefs Brand Radar at $20 per 1,000 on its custom prompts path returns the aspect and the sentence behind each mention, and reading custom prompt data back costs no API units. Its terms license the data for your internal business operations only, so an agency report needs another route.
  3. You already collect the answers. Score them yourself with a cheap model and a written rubric, at about $0.09 per 1,000 answers in tokens, and spend the saved subscription on the human sample that keeps the classifier honest.

Whichever way you go, collapse the headline to one question before the first report: what share of the answers that name us say something negative, and about what. That number survives a change of vendor, a change of scale and a change of rubric, which no positivity score does. Our walkthrough of tracking a brand in ChatGPT with an API covers the mention side of the same pipeline, and tracking the citations of your pages covers the sources that produced the framing in the first place.

Frequently asked questions

What is an AI brand sentiment API?

It is an API that runs your prompts on the AI engines, keeps the answers, and attaches a polarity to each mention of your brand: a label such as positive, neutral, negative or mixed, a score, or both. No engine API returns it, so the figure is always computed by a measurement vendor or by you over stored answers. Rates published on October 23, 2026 for collecting and scoring 1,000 answers run from $2.94 on Rankability's $97 Starter plan and $3.30 on SE Visible's $99 Basic plan to $20 on Ahrefs Brand Radar's custom prompts path and $81.94 on AthenaHQ's $295 Starter plan.

Which APIs return sentiment per mention rather than per answer?

Three of the ones we track. Ahrefs Brand Radar's Get AI responses endpoint returns a `brand_sentiment` array with one entry per mention, each carrying the brand, a sentiment among positive, neutral, negative and mixed, an `aspect` among eleven values such as `price_value` and `service_support`, and the sentence around the mention. Conductor's `ai_search_sentiments` endpoint states that "each row represents a single sentiment mention for a brand or product within an AI response", with a `sentiment_category` among nine values. Promptwatch returns a `brandMentions[]` array where each brand carries its own `sentiment`, `position`, `visibilityScore` and `toneOfVoice`. Scrunch and OtterlyAI work one level up, per response and per brand aggregate.

How much does it cost to measure brand sentiment on 1,000 AI answers?

At the prices published on October 23, 2026: $2.94 with Rankability, $3.30 with SE Visible, $13 with API for GEO down to $7 at its largest tier, $13.32 with GEOly, $15.75 with OtterlyAI, $15.83 with Promptwatch, $20 with Ahrefs Brand Radar on its custom prompts path, €30.67 with LLM Pulse and $81.94 with AthenaHQ, with Profound, Peec AI, Scrunch and Conductor selling API access inside a subscription or on quote. API for GEO, our pick after hands-on testing by Mathias C., bills one credit per prompt, per model and per run, states that reads over stored runs are not billed, and allows resale in writing.

Can I compare a sentiment score from two different tools?

No, because the scales are built differently. Peec AI publishes its aggregation as `((sum(sentiment_sum) / sum(sentiment_count)) / 2 + 0.5) * 100` on a 0 to 100 range, and notes that most scores land between 65 and 85. Scrunch scores each qualifying response positive 100, mixed 50, negative 0 and averages them, warning that 85 "does not mean '85% positive'". OtterlyAI returns counts of `positive`, `neutral` and `negative` plus an `nss`. Ahrefs returns labels with no score at all. Two vendors can report the same month as 68.8 and as +37.5 without either being wrong.

How accurate is an automated sentiment label?

Less accurate than the single number suggests, and the ceiling is human disagreement. A study of six human raters labelling 100 tweets, published on arXiv on October 7, 2026 and accepted at TPDL 2026, measured a Fleiss' kappa of 0.39 between the humans on three-class sentiment, rising to 0.45 when the task was collapsed to negative against non-negative. Against the human majority vote, the three open models tested scored between 0.48 and 0.63 on three classes, while a sentiment model fine-tuned on tweets reached 0.73 and 0.88 on the two-class split. The texts were tweets rather than AI answers, so treat the figures as an order of magnitude, and keep the snippet behind every label so a human can audit it.

How many answers do I need before a sentiment figure is stable?

Several hundred per period. cloro's published study of August 14, 2026 measured a mean overlap of 0.40 between the cited domains of two runs of the same prompt on the same model minutes apart, puts about 300 answers per period at roughly ±5 percentage points, and needs about 294 answers on each side to support a claimed move from 20% to 30%. A sentiment figure is narrower than that, since it is computed only over the answers that name your brand: on 800 collected answers with your brand in 38% of them, the score rests on about 304 labels.