Managing Sentiment in AI Recommendations: From Neutral to Advocate

1. Why “neutral” is a competitive loss, not a safe place

In the search era, a neutral mention was fine. A directory listing, a passing reference, a name in a roundup — all of it added a little authority and did no harm. In the recommendation era, neutral is where deals quietly die. When a buyer asks an assistant “which [category] tool should we choose?” and the model lists you flatly between two rivals it describes warmly, you have not been attacked; you have been out-framed. The buyer reads the enthusiasm gap and follows it. Nothing false was said. You simply lost on tone.

This is the uncomfortable truth of sentiment in AI answers: the model is not just deciding whether to name you, it is deciding how to characterise you, and that characterisation does more conversion work than the bare mention ever did. A warm “widely regarded as the best fit for UK SMEs” is a different commercial event from a clipped “also available.” Both name you. Only one recommends you. The brands that treat AI sentiment as a managed asset — rather than an accident of whatever the models happen to have read — are the ones that will convert the answer layer instead of merely appearing in it.

You have not been attacked. You have been out-framed — and the buyer follows the enthusiasm gap.

The shift is sharpest in UK B2B, where buyers increasingly open a vendor shortlist by asking an assistant rather than a search box. By the time a procurement lead reaches your website, the model has often already framed you against two rivals — and that framing, not your homepage, set the tone for the whole evaluation. A British buyer is also reading a British-inflected answer: models pick up the dry, understated register of UK source material, so a lukewarm UK trade-press line (“does the job”) can land as a colder signal than its American equivalent. Managing sentiment is, in effect, managing the first impression that now happens entirely outside your analytics.

The goal of this edition is to take you up a ladder: from being characterised as a detractor would, to neutral presence, to active advocacy in which the models recommend you in your own words and on your own terms. Crucially, this is not about manipulating models or gaming reviews — in the UK, much of that is now explicitly unlawful, as Section 7 sets out. It is about systematically improving the genuine signals the models read, so that an honest model, reading an honest record, reaches a warmer conclusion.

2. What sentiment means inside an AI recommendation

AI-answer sentiment is not the same thing as social-media sentiment, and conflating them is the first mistake. Social sentiment measures how people feel in posts you can find. AI-answer sentiment measures how a model characterises you in a synthesised reply — a blend of the source sentiment it has absorbed, the comparative framing it constructs, and the qualifiers it chooses to attach. You can have glowing social sentiment and still be framed coolly by a model that leant on one influential, lukewarm review. The two correlate, but they are not the same surface, and only one of them is now shaping the shortlist.

Sentiment in an answer expresses itself through five observable signals, and reading them precisely is what separates a real sentiment programme from vague “the AI doesn’t like us” anxiety:

• Recommendation strength — are you actively recommended, merely listed, or listed last? This is the dominant signal.
• Qualifiers — the adjectives and caveats attached: “premium” vs “expensive”, “powerful” vs “steep learning curve”, “UK-focused” vs “UK-only”.
• Comparative placement — where you sit relative to named rivals, and whether the contrast flatters or diminishes you.
• Framing of weaknesses — whether a genuine limitation is stated neutrally, sympathetically, or as a reason to avoid you.
• Hedging — uncertainty language (“some users report…”, “it may be…”) that signals the model is unsure and so under-commits to recommending you.

The table below maps these into a working five-point sentiment scale, which the rest of the edition uses for scoring. The point of a fixed scale is comparability: two analysts, and your monitoring script, should land on the same band for the same answer.

One distinction underpins everything that follows: sentiment is not the same as accuracy, and the two demand opposite responses. An answer can be entirely factual and still cost you the deal because it is framed coldly — “capable but expensive, and the interface takes some learning” may be true in every word and still steer a buyer to a warmer-sounding rival. That is a sentiment problem, addressed by improving genuine signal, never by disputing facts. An answer can also be warm and wrong, which is a different and more urgent problem. Throughout this edition, when a negative framing is true you manage it; when it is false you correct it. Mixing the two up is the most common and most damaging mistake in the discipline.

3. The Sentiment Ladder and Maturity Model (your deliverable)

Here is the framework, placed before any tooling or measurement detail so it is usable immediately. Managing sentiment is the disciplined work of climbing two ladders in parallel: moving individual answers up the sentiment scale, and moving your whole organisation up a maturity curve from reactive to systematic. Take both together.

The Sentiment Ladder: the climb you are actually making

Every brand sits somewhere on the five-band scale above for any given prompt, and the job is to climb it one rung at a time. The mistake is trying to leap from Detractor to Advocate in a single campaign. Sentiment moves rung by rung as the underlying signals shift, and each rung has a characteristic lever:

The Sentiment Maturity Model: where your organisation sits

The second ladder is organisational. Most UK brands are at Stage 1 or 2 — they have never read what the models say about them, or they have read it once and panicked. The destination is Stage 5, where sentiment is a tracked metric with an owner and a routine. Locate yourself honestly; you cannot skip stages.

4. Where AI sentiment actually comes from

You cannot manage sentiment by talking to the model; you manage it by changing what the model reads. AI-answer sentiment is assembled from inputs, and almost all of them are signals you can legitimately influence. Understanding the inputs is what turns sentiment from a mood into a lever.

The dominant input is third-party source sentiment: how the pages, reviews and references the model trusts characterise you. Models do not invent an opinion of your brand from nothing; they aggregate the tone of what authoritative sources say, weighting recent and well-corroborated sources heavily. This is why the mechanics of how the major assistants select and weight product sources — covered in how ChatGPT, Perplexity and Gemini choose what to recommend — are the foundation under any sentiment work. If you do not know which sources a model leans on, you are managing sentiment blind.

The second input is review-corpus sentiment. For most consumer and many B2B brands, aggregated review platforms carry disproportionate weight, because they are structured, plentiful and explicitly evaluative. A model reading thousands of recent four- and five-star reviews with specific praise builds a warm, confident picture; a model reading a thin, stale or mixed review corpus hedges. Review recency and specificity matter as much as the average score — a wall of vague five-stars persuades a model less than a steady stream of detailed, recent, plausibly-genuine ones.

The third input is comparative framing in buying-guide content. When the model constructs a comparison, it leans on the comparison content that already exists — roundups, “best-of” guides, head-to-head articles. If the guides the model trusts frame you as “premium but pricey,” that framing propagates into answers. Earning a place in, and shaping the tone of, the comparison content models cite is direct sentiment work — the practical how-to lives in earning citations in AI buying guides.

The fourth input is structured and first-party signal: your own clearly-stated, well-structured information about what you do, who you serve and what you are good at. Models lean on a brand’s canonical self-description when third-party signal is thin or contradictory. Ambiguous, hedged or absent first-party framing leaves the model to infer your positioning from whatever else it finds — often a competitor’s comparison page. Saying clearly and consistently what you are best at is not spin; it is giving an honest model an honest, structured answer to characterise you with.

Cutting across all four inputs is recency and weighting, the variable most brands underestimate. Models do not treat every source equally or timelessly; they lean hardest on recent, well-corroborated, high-trust sources, which means an old but influential piece can anchor your sentiment long after reality has moved on, and a single authoritative recent review can outweigh a hundred older ones. This is why sentiment is rarely fixed by volume alone. A brand that redesigned its product two years ago but whose most-cited reviews predate the change will keep being framed by the old reality until fresh, specific, trusted signal accumulates enough weight to displace it. Recency is not a tie-breaker; it is often the whole game.

You cannot manage sentiment by talking to the model. You manage it by changing what the model reads.

5. Measuring sentiment at scale (a working approach)

At small volume you score sentiment by hand against the five-band scale. To track it over time across a full prompt panel and several models, you classify the answers you have already captured. The pattern extends the monitoring loop: take the stored answers, pass each to a classifier with a fixed rubric, and write back a sentiment band plus the qualifiers found. The snippet below is illustrative — a minimal classifier you adapt, not a product — and deliberately uses a fixed rubric so results are reproducible rather than vibes.

The qualifier extraction is the quiet workhorse here. The band tells you where you sit; the harvested qualifiers tell you why, and they aggregate into the single most actionable artefact a sentiment programme produces — a ranked list of the caveats the models keep attaching to you. That list is what Section 6 acts on.

The cost maths at volume

Classification cost scales with the number of answers that name you, not the whole panel. Multiply it out: mentioned-answers per cycle × cycles per month = classification calls per month. If a 30-prompt panel run five times across three models names you in, say, 40% of its ~450 answers per weekly cycle, that is ~180 classification calls a cycle, ~775 a month. As a plug-in formula, per-call cost ≈ (in_tokens × I + out_tokens × O) / 1,000,000, where I and O are your model’s per-million input/output rates; classification calls are short in and short out, so on a small model the monthly cost is trivial — single-digit dollars territory. Engineering and review time dominate, not tokens.

Where this breaks in production

Automated sentiment classification fails in predictable ways, and a programme that does not plan for them will quietly report nonsense. Mixed-sentiment answers (“powerful but pricey”) force a single band onto a genuinely two-sided characterisation — capture the qualifiers, not just the band, so the nuance survives. Sarcasm and British understatement fool classifiers regularly, which is a real UK hazard given how much native source material relies on dry, indirect tone. Classifier self-inconsistency means the same answer can score differently on re-runs, so treat any single score as noisy and trust the distribution. And the highest-stakes failure is the false detractor: a true-but-unflattering statement wrongly escalated as something to “fix,” when the honest move is to improve the product, not the framing.

Failure thresholds and the cheaper fallback

Set two thresholds. A confidence threshold: if classifier output fails to parse or the model hedges its own scoring on more than a set share of calls, flag the cycle as low-confidence rather than trusting it. And a human-review threshold: any answer scored Detractor (−2), and any new qualifier appearing for the first time, gets a human eye before it triggers action. The cheaper fallback when budget or volume bites is to batch several answers into one classification call (cutting calls sharply at a small accuracy cost), or to drop back to a keyword-based first pass — scanning stored answers for a watch-list of known qualifiers (“expensive”, “difficult”, “limited”) — which is near-free and still surfaces the recurring caveats that matter most.

Reproducibility metadata

Every sentiment cycle should record the rubric version, the classifier model and version, the capture and classification dates in UTC, the band scale used, and whether classification was human, model-assisted or fully automated. Without the rubric version in particular, a shift in your sentiment trend is uninterpretable — it could be a real change in how models see you, or simply you having quietly reworded the rubric. Log the conditions and the trend is evidence; omit them and it is decoration.

6. The recurring-qualifier playbook: turning caveats into advocacy

The ranked qualifier list from Section 5 is where sentiment management becomes concrete. A qualifier that appears once is noise; a qualifier that recurs across models and cycles is a narrative the models have settled on, and it almost always traces to a small number of influential sources. The playbook is a loop: identify the recurring qualifier, trace it to source, decide whether it is true, and act accordingly.

The decision fork matters more than anything else here, because it determines whether you are doing legitimate sentiment work or trying to paper over a real flaw. If the qualifier is false — the model says “UK-only” when you serve the EU, or “no integrations” when you have dozens — it is a factual error, and you route it into correction rather than sentiment work. If the qualifier is true but outdated — “steep learning curve” from reviews predating a redesign — the lever is recency: generate fresh, accurate signal so the model’s picture updates. And if the qualifier is true and current — you genuinely are the expensive option — then no amount of source-tinkering is legitimate or effective; the move is either to fix the underlying issue or to reframe it honestly (“premium” is the same fact as “expensive,” told by a source that explains the value).

This is also the boundary line of the whole discipline. Everything above the line — fixing errors, refreshing stale signal, earning honest corroboration — is legitimate and durable. Everything below it — fabricating reviews, planting fake testimonials, manufacturing consensus — is both unlawful in the UK and self-defeating, because models increasingly discount low-trust, low-diversity signal. When a recurring qualifier reflects a competitor actively seeding misinformation rather than an honest reading, that is a different problem with its own response, covered in our forthcoming guide to competitor misinformation in AI answers.

A worked example makes the loop concrete. Consider a Manchester-based B2B software firm whose monitoring showed a healthy Share of Model Voice but a sentiment average stuck around Neutral, with one qualifier — “steep learning curve” — recurring across three assistants over six weeks. Tracing it to source revealed the framing came from a cluster of detailed reviews written before a major onboarding redesign eighteen months earlier; the caveat was true once and outdated now. The legitimate move was not to dispute the old reviews but to generate fresh, specific, recent signal: prompting genuinely satisfied recent customers to describe their onboarding experience, and earning an updated walkthrough in a trade outlet the models cited. Within a couple of monitoring cycles the qualifier’s frequency fell and the sentiment average drifted toward Positive — not because anything was manipulated, but because the model’s evidence base had been honestly refreshed.

7. The UK rules that decide which levers you may pull

Sentiment management is more legally constrained in the UK than in most markets, and the constraints are not abstract — they rule specific tactics in or out. This is the section to read before you brief anyone to “improve our reviews.”

The fake-review ban is now black-letter law

As of 2025, the Digital Markets, Competition and Consumers Act 2024 makes it explicitly unlawful to write, commission, host or incentivise fake or concealed-incentive reviews, with the Competition and Markets Authority empowered to levy substantial fines. This is the single most important fact in UK AI-sentiment work, because review-corpus sentiment is a primary model input (Section 4) and the obvious-but-illegal shortcut is to manufacture it. You cannot. Genuine review generation — asking real customers, at the right moment, without paying for a particular score — is legitimate and effective; anything that fabricates or hides incentives is now a regulatory risk that dwarfs any sentiment upside.

FCA: you cannot manufacture sentiment you are not allowed to claim

For regulated UK financial firms, sentiment management collides with the Financial Conduct Authority’s requirement that communications be fair, clear and not misleading. You cannot seed sources with favourable framing that overstates performance, downplays risk, or implies an endorsement you do not have — even indirectly, even through third parties, and even if the goal is only to warm up an AI answer. The fair-clear-not-misleading test follows the claim wherever it surfaces. For regulated firms, the safe and effective sentiment lever is accurate, compliant, well-structured first-party clarity plus genuine customer signal — never engineered framing.

ASA and GDPR: substantiation and personal data

Two further constraints bind everyone. The Advertising Standards Authority requires that marketing claims be substantiated and that testimonials be genuine and held on file — so any positive framing you originate, and that then propagates into AI answers, must be defensible. And UK GDPR governs the review, testimonial and sentiment data you collect and store: customer reviews naming individuals, captured AI answers mentioning real people, and testimonial records all constitute personal data needing a lawful basis, sensible retention and security. The Information Commissioner’s Office treats AI-related processing as fully in scope. None of this blocks legitimate sentiment work; it simply means the legitimate path — genuine, substantiated, well-governed signal — is also the only lawful one.

8. The advocacy flywheel: compounding from neutral to advocate

The top rung — Advocate — is not reached by a campaign; it is reached by a flywheel that, once turning, defends itself. Each genuine positive signal makes the next one easier, and the models reward the density and consistency that result. The loop has four stages, each feeding the next.

1. Earn genuine corroboration. Real customer reviews, honest third-party coverage, and accurate buying-guide placements create warm, well-sourced signal in the places models trust.
2. Models recommend more warmly. Denser, more consistent positive signal moves you up the sentiment scale and makes the model more confident in recommending you actively rather than hedging.
3. Warmer recommendations drive more, better customers. Buyers who arrive via a confident AI recommendation are better-qualified and more satisfied — and more likely to leave the kind of specific, genuine review that feeds stage one.
4. The signal compounds. A steady flow of recent, specific, genuine corroboration is exactly what a model weights most heavily, so each turn of the loop raises the floor under your sentiment.

The flywheel is why sentiment is a moat rather than a campaign. A competitor can copy your messaging overnight; they cannot copy years of accumulated genuine corroboration. This is also where sentiment management connects to the wider recommendation strategy — the same dense, trusted signal that warms your sentiment is what gets you recommended at all, the foundation laid out in our hub on getting recommended by AI shopping agents. Sentiment and recommendation are two readings of the same underlying asset.

Measure the flywheel, or you will not know it is turning. The leading indicator to watch is Advocacy Rate — the share of your category and comparison answers that land in the Positive or Advocate bands — read as a trend alongside the falling frequency of your worst qualifiers. A rising Advocacy Rate with a shrinking qualifier list is the signature of a working programme; a flat Advocacy Rate despite heavy activity usually means you are adding signal in places the models do not weight, or adding volume where they want recency and specificity. Annotate the trend with what changed each cycle — a review push, a trade placement, a product fix — so the line becomes a story you can repeat deliberately rather than a number you hope keeps climbing.

9. Signal-to-play decision matrix

To keep the programme operational rather than analytical, route each sentiment signal to a predetermined owner and play. The matrix below turns the metrics into a standing response system.

False-claim signals branch out of this matrix into correction rather than sentiment work — you cannot reframe a falsehood, you fix it — using the hallucination-correction playbook. Keep the two workflows distinct: sentiment management shapes how a true picture is framed; correction repairs a picture that is false. Confusing them wastes effort and, in regulated UK sectors, can create the very compliance risk you were trying to avoid.

10. Your 90-day advocacy roadmap

Climb deliberately. Each phase produces something usable, so value compounds rather than waiting on a finished system.

• Days 1–30 — Baseline and diagnose. Lock a buyer-prompt panel, capture and score answers on the five-band scale, and build the ranked qualifier list. You now know your sentiment average, your worst prompts, and the caveats the models keep attaching. Locate yourself on the maturity model.
• Days 31–60 — Trace and fix. Run each recurring qualifier through the Section 6 fork: route false ones to correction, refresh outdated ones with genuine recent signal, and decide honestly on the true-and-current ones. Begin genuine, lawful review generation where your buyers and the models actually look.
• Days 61–90 — Compound and defend. Stand up the flywheel: a repeatable cadence of genuine corroboration and comparison-content placement, with sentiment re-scored each cycle to confirm the climb. Install the signal-to-play matrix and a fortnightly review so sentiment becomes a managed, owned metric — Stage 5.

11. What to do on Monday

Neutral is not safe; it is a slow loss to whichever competitor the model describes more warmly. But the path off the neutral rung is not manipulation — in the UK it cannot be — it is the patient, lawful work of improving the genuine signals an honest model reads. Measure how you are framed. Harvest the qualifiers. Trace them to source. Fix what is false, refresh what is stale, own what is true, and earn the corroboration that warms the rest.

So on Monday, do the small version. Score your ten most important buyer prompts on the five-band scale, write down every qualifier attached to your name, and sort them into false, outdated, and true. That sorted list is your sentiment strategy on a single page — and somewhere on it is the caveat that is costing you the most deals right now. Start with that one. Everything else in this special edition is the system that turns that one page into a flywheel.