How AI Talks About Your Brand: Sentiment as a Compounding Asset

Most visibility dashboards answer one question: does the model mention us? The harder question is what it says while mentioning us — because in an answer engine, characterisation travels with the recommendation. “A solid budget option with mixed support reviews” and “the category standard” are both mentions. Only one of them sells.

Chris Donnelly’s framing at Searchable makes the structural point: sentiment in this environment compounds, because models re-ingest text that earlier model output helped shape. A search ranking resets when the algorithm updates; a characterisation embedded across sources persists until the sources change.

This is the tenth report in our research series, following engine divergence, the citation supply chain, brand misinformation, the B2B shortlist, the long-tail thesis, enterprise agents, challenger versus incumbent, the shopping source map and general versus vertical assistants.

Methodology

A synthesis, openly sourced

This report is a research synthesis, not a Brantial panel measurement. It explains the mechanism of sentiment formation from published research and 2026 conference presentations, cited inline; interpretation is labelled as interpretation. It deliberately publishes no brand-level sentiment scores: those require first-party measurement, and the Turkish sentiment map is planned as a follow-up on Brantial’s prompt panel.

The frame

Four questions: where sentiment comes from, why it hardens, how it differs from a wrong fact, and what a brand can move.

Finding 1: Sentiment is assembled from sources, not authored by the model

A model does not form an opinion about a brand; it summarises the balance of what its sources say. That makes the source mix the sentiment mix — and our source ecosystem report established what that mix looks like: fewer than 3% of citations to tier-1 media, the rest to niche, specialist and community sources (Tim Sanders/G2; Josh Blyskal/Profound — Zero Click NY 2026).

The engine split matters more here than anywhere else. ChatGPT lifts community content into roughly 16% of citations; Claude sits at 0.9% (Profound). Community text is where evaluative language lives — praise, complaint, comparison, resentment. An engine weighted toward communities inherits their tone; an engine weighted toward structured editorial inherits a flatter, more specification-driven characterisation of the same brand.

Interpretation: a brand can carry materially different sentiment on two engines without anything about the brand differing — only the sources being read.

Finding 2: The loop is what makes it compound

Three published observations describe a closing loop.

Consumers verify in the same places models read. Around half of consumers check an AI recommendation on Reddit before acting (Rob Gaige, Reddit — Zero Click NY 2026). The verification then becomes another thread — new text, in a source the model already trusts.

Brand intrusion degrades the source. Reddit’s own data: community sentiment collapses for brands posting more than three times a week (Gaige). The correction attempt itself can worsen the corpus the model reads.

Model output re-enters the corpus. Donnelly’s compounding argument: as AI-assisted text spreads across the web, later models read characterisations that earlier models helped produce.

Interpretation, labelled as such: the compounding claim is a mechanism argument, not a measured coefficient. No published study quantifies how fast a characterisation hardens. What the evidence supports is direction, not rate — which is enough to justify measuring early rather than reacting late.

Finding 3: A wrong fact and a bad characterisation need different work

Our misinformation report covered claims that are simply false: an outdated price, a discontinued product, the wrong ownership. Those are correctable at the source, and re-verifiable afterwards.

Sentiment is not that. “Support is slow” may be an accurate summary of what many sources say. Correcting it is not an editing problem but an evidence problem: the balance of published experience has to change before the summary changes. Two implications follow.

Facts have a fix; characterisations have a lag. A corrected specification can propagate within a crawl cycle. A characterisation shifts only as new sources accumulate — slower, and largely outside a brand’s direct control.

Suppression is not available. In a citation-driven system, the strategy of drowning criticism does not work: the sources remain readable, and the attempt itself is visible in the corpus.

Finding 4: The four measurable questions

Reading the published work together, four sentiment questions are answerable today with existing measurement — and they are the ones worth instrumenting before any campaign.

1. What adjectives travel with the brand? The recurring evaluative language in answers, per engine.
2. Which sources supply them? The specific pages and threads a model cites when the characterisation appears — the actionable layer.
3. Where does it diverge by engine? Given the 16%/0.9% community split, comparing engines separates “our reputation” from “the corpus one engine happens to read”.
4. How does it move after a change? Re-measuring after a source-level fix is the only way to distinguish a durable shift from noise, since answers are non-deterministic.

Donnelly’s four questions for brands — mention frequency, sentiment, what AI says about you, and which sources feed it — map onto the same instrumentation, and his blunter point stands: the claim that attribution here is impossible is a myth.

What this means in Türkiye

The community layer is complaint-weighted. Turkish commercial discussion concentrates on Ekşi Sözlük, ŞikayetVar and category forums. ŞikayetVar in particular is structurally complaint-oriented: a brand’s presence there is, by design, its unresolved problems. If the ChatGPT community-citation pattern transfers to Turkish, this is a sentiment exposure most brands have never quantified — and the first thing worth measuring.

Thin corpora move faster in both directions. With fewer Turkish sources per category, individual threads carry more weight. That cuts both ways: a single well-argued negative thread can dominate a characterisation, and a small number of substantive, credible sources can shift it faster than in English.

Turkish-language measurement is not optional. A brand measuring sentiment only in English is measuring a different corpus than the one its domestic buyers’ prompts reach. The two can disagree entirely, which is precisely why the follow-up study measures Turkish prompts separately.

The takeaway

Sentiment in answer engines is a summary of the sources, hardened by a loop that runs through the communities buyers use to check AI recommendations. It is not editable the way a fact is: the work is source-level, slow, and mostly indirect — earn better-evidenced coverage, resolve the complaints that generate the language, and stay out of the posting behaviour that degrades the corpus further.

The sequence: instrument the four questions above per engine, separate wrong facts (fixable now) from characterisations (shiftable over time), and re-measure after each source-level change rather than after each campaign.

Sources

• Chris Donnelly, Searchable — compounding sentiment, the four brand questions and attribution, 2026
• Rob Gaige, Reddit — verification behaviour and brand-participation limits, Zero Click NY 2026
• Josh Blyskal, Profound — citation distribution by source type and engine, Zero Click NY 2026
• Tim Sanders, G2 — citation ecosystems and tier-1 media share, Zero Click NY 2026
• Brantial research series (2026): Source Ecosystem, Brand Misinformation, Divergence

Figures are as presented in the cited talks and publications and primarily describe English-language, US-market behaviour. No brand-level sentiment scores are published here; Brantial’s follow-up study will map Turkish brand sentiment on our own prompt panel.