Recommended for the right reasons

Sentiment & Recommendation Analysis

A clear read on how models describe the brand, why they recommend it, and when a recommendation is flattering but wrong.

Repeated recommendation scenarios reveal tone, shortlist position, stated reasons, caveats, and non-fit cases. The result is useful when those reasons can be checked against documented buyer criteria and evidence. Product and content leads who need to know whether the brand is being recommended for reasons its evidence supports.

You can compare the stated recommendation reasons with approved buyer criteria and evidence, including cases where the brand should remain out of the shortlist.

Figure reading positive, neutral, and negative tone gauges beside an AI answer card

Some of the 500+ brands we've worked with

See all references
  • Sabancı Üniversitesi
  • MNG Kargo
  • Obilet
  • Joker
  • Bluemint
  • S Sport Plus
  • Vispera

From buyer criteria to fit tests

From buyer criteria to fit and non-fit tests

Documented decision criteria come first. Eligible and ineligible scenarios then run under fixed conditions, and any gain that creates more wrong-fit recommendations fails the test.

How we hold ourselves to it

  • Fit checked before praise
  • Reasons traced to evidence
  • Criticism isn't a defect
  • No tone dialed to order

  1. Extract decision criteria from audience questions

    Sales, support, and research questions are grouped by job, constraint, risk, and stage. Uncommon but material scenarios remain in the set.

    Recommendation intent matrix with persona, need, constraints and expected decision.

    AI assist
    Extracts buyer needs, exclusions, conditions and decision criteria from approved audience research without inventing personas or demand.
    Human gate
    Client research, sales and product owners approve criterion provenance, priority and intended recommendation context.

  2. Define fit, non-fit and evidence rules

    Each criterion gets a support rule, a disqualifier, and an acceptable evidence standard before brand appearance is observed.

    Criteria-to-evidence ledger with fit, non-fit and uncertain states.

    AI assist
    Maps each criterion to approved capabilities, disqualifiers, comparative evidence and an explicit uncertain state.
    Human gate
    Domain, legal and product owners approve fit, non-fit and evidence rules before scenarios run.

  3. Run repeated recommendation scenarios

    Fixed eligible, ineligible, ambiguous, and competitor scenarios run repeatedly, preserving each shortlist and its stated reasoning.

    Fit and non-fit decision table linked to complete answer captures.

    AI assist
    Executes the fixed eligible, non-fit and ambiguous scenarios and preserves full shortlists, reasons, caveats, sources and failures.
    Human gate
    A monitoring operator validates run completeness and applies the same validity rule to every scenario.

  4. Audit recommendation reasoning and sources

    Every inclusion, omission, and caveat is compared with the approved criteria. Exposed sources are checked against the reason the model gave.

    Annotated shortlist evidence set with support and harm labels.

    AI assist
    Maps every recommendation reason to exact evidence and classifies supported, partial, unsupported, contradictory and harmful outcomes.
    Human gate
    Domain, editorial and legal reviewers approve evidence and harm labels or record an abstention.

  5. Close honest evidence and content gaps

    Verifiable criteria pages, comparisons, limitations, or disclosures are strengthened only where the brand can substantiate the underlying fit.

    Prioritized buyer-criteria content specification and evidence backlog.

    AI assist
    Drafts bounded fact, comparison, fit and exclusion improvements from approved evidence and identifies gaps that must remain unanswered.
    Human gate
    Content, product, brand and legal owners approve truthful corrections within their authority. No false-fit claim may ship.

  6. Retest paraphrases and unseen personas

    The original panel and unseen persona or wording variants run again to check whether the changes introduce recommendations for non-fit cases.

    Retest report covering eligible and ineligible precision.

    AI assist
    Replays paraphrases, eligible cases, non-fit cases and holdout personas and reports supported recommendations and false positives separately.
    Human gate
    An independent reviewer rejects any gain that increases unsupported or wrong-fit recommendations.

The recommendation map you get

A recommendation map that values fit over flattering mentions

The handoff makes supported fit, non-fit cases, missing evidence, and unreliable recommendation reasons visible enough for product, content, and legal owners to decide.

  • Recommendation trigger baseline

    A scenario-level baseline of eligible recommendations, false positives, omissions, rationale accuracy and ordinary variance.

    Accepted when

    Reveals whether the priority problem is missing eligibility evidence, incorrect fit or unstable observation.

    Cadence: Variance separated

  • Buyer-criteria content specification

    Defines the questions, proof, limitations and comparison language each priority page must make inspectable.

    Accepted when

    Tells content owners what to publish without writing unsupported superiority claims.

    Cadence: Evidence required

  • Evidence-gap and exclusion roadmap

    Separates proof the team can create, evidence requiring independent validation and contexts the brand should explicitly decline.

    Accepted when

    Supports investment, disclosure, deferment or a deliberate non-action for each criterion.

    Cadence: Gaps separated

  • Retest report on unseen personas

    Shows how the original and unseen personas changed, including new non-fit recommendations and unsupported reasons.

    Accepted when

    Determines whether the release improved recommendation correctness, with raw mention count reported separately.

    Cadence: Retested on unseen cases

  • Eligible recommendation rate

    Counts recommendations only where the approved fit rules are met, with raw mentions and ineligible recommendations excluded.

    Accepted when

    Documented measurement rule: fit-qualified runs that explicitly recommend the brand. The scope, the runs it is counted against, the exclusions and the time window are all fixed before any comparison.

    Cadence: Rule fixed upfront

When the reasons are wrong

When the brand appears for the wrong reasons, or not at all

This method applies when recommendation answers omit the brand, place it with a poor-fit audience, or give reasons the product cannot honestly support.

A good fit when

  • Relevant shortlists omit the brand or misstate its fit — The brand may be absent from ChatGPT shortlists, or included for unsupported reasons or audiences.
  • A decision is waiting on fit criteria — The team needs to know which criteria the brand meets, where proof is missing and which contexts are unethical.
  • Buyer proof and recommendations are split — The team has research, product evidence and ChatGPT examples, but no shared rule for judging recommendation fit.
  • Buyer criteria define recommendation eligibility — We turn approved questions into use case, budget, location, risk and must-have conditions for each recommendation.
  • Brand fit and exclusions stay evidence-backed — We record where the offer fits, where it doesn't, and which approved proof supports each comparison criterion.
  • Recommendation reasons and shortlist position — We inspect why repeated shortlists include, omit or qualify the brand. A bare mention isn't a correct recommendation.
  • Reviews and proof stay traceable — Reviews, comparisons, credentials and outcome claims stay attributable, current and disclosed. We do not manufacture endorsements.

Better handled as other work when

  • You want tone engineered to order — We can report how models currently characterise you and why. We cannot dial sentiment to a target figure on request.
  • Recommendations remain outside our control — We don't manipulate hidden prompts, fabricate reviews or manufacture consensus. Third parties make the recommendation.

People who watch how AI cites a brand

GEO work starts with recording what AI answers actually say about a brand today, then moves to the parts you can influence. The consultants below work on the specific capability this page covers.

Content we've produced on this topic

Articleby Can MutioğluGenerative AI Optimization (GEO) for E-Commerce SitesArticleby Didem HimmetliThe Transformational LLMs on Search and Recommendations

Tools we use

Tools behind this work

  • Peec AI

    tracks sentiment as its own metric, separate from whether the brand was simply mentioned

  • Otterly.AI

    connects a recommendation pattern back to the specific content gap causing it

  • Airtable

    holds the documented buyer criteria a recommendation's stated reasons get checked against

A non-fit is a correct recommendation too

Check why a model recommends the brand

One buyer question, its fit rules, and the supporting proof define the test case. We compare sampled recommendations and label each context as supported, uncertain, or out of scope.
Review recommendation fit

Questions we hear

Recommendations remain a third-party decision. We don't use hidden prompts or planted reviews. Manufactured consensus also sits outside the method. No recommendation or citation is guaranteed.