GEO-119·Issue 10·Authority & Off-site·119 / 132
Negative Mentions in AI Answers: Monitor and Repair
Negative AI answers usually quote an old thread, a stale review, or the wrong entity, not today's product. A triage table, four repair paths, and a re-test.
Issue focusMeasurable: AI Visibility, Attribution, and Experiments
A composite case: three weeks before Black Friday, the growth lead at a US brand selling app-controlled irrigation controllers opens a screenshot. A shopper asked ChatGPT whether the controllers are reliable; the answer leads with “the app drops schedules and support is slow.” Firmware was rebuilt in 2024; the answer still quotes a 2023 Reddit thread and a review site that stopped testing the category in 2022.
Negative mentions in AI answers are a corpus problem before a PR one: the model repeats the most detailed old sentence it can retrieve, so classify each claim by source, repair the right layer, and re-test.
Triage first: three sources, three repairs
Treating negative mentions as one blob is the expensive mistake: “the app drops the schedule” needs a different owner per source.
| Source | Typical shape | Why the model keeps it | What you can move |
|---|---|---|---|
| Old news and older reviews | A dated recall post; a test of a discontinued model | Dated, detailed, still retrievable, widely linked | Medium: add a newer record; the original stays |
| Marketplace complaints | One-star reviews, shipping complaints, agency filings | Specific, emotional, third-party | Low: you can only shift the ratio |
| Wrong entity | A same-name brand, a retired category, a mixed-up founding year | Entity resolution chose the wrong node | High: site, directories, and knowledge bases |
A genuinely broken product needs a product fix, not a table row: GEO decides how others see a problem, not whether it exists.
Why the model prefers the old complaint
Complaints travel better than praise, and the reason is citability, not bias.
A five-star review says “great build, works perfectly” — no condition, date, or number a model can check, so it is skipped. A complaint carries scene and quantity: “third charge, still wouldn’t pair, refund took nineteen days.” Granular sentences get lifted whole; old threads have replies and links keeping them near the top.
Your site argues; a thread asserts. “New connection stack, redesigned in 2024” is a claim a model must vouch for; “bought in August 2023, the app kept dropping” is a dated fact. The answer keeps the specific one, then appends yours as “the company says” — the averaging that brand narrative consistency audits across fields.
There is a structural reason too: repairs happen on your domain, the negative off-site, and your site holds a minority of retrieval slots. A new page changes nothing until a dated off-site record competes with the thread.
Five steps from monitoring to re-test
Skip the new tool: add one field — negative claim — beside the AI visibility metrics you track.
| Step | Action | Owner | Passes when |
|---|---|---|---|
| 1 Extend prompts | Six types: drawbacks, complaints, returns, support, recalls, comparisons | Growth, regional ops | Each prompt has a market, language, owner |
| 2 Capture sentence and source | Exact wording, cited URL, date, region; no paraphrase | Content | Every claim maps to a source URL |
| 3 Classify | Factual error, stale experience, or real defect, one owner each | Brand, category, support | No “pending” bucket; defects route to product |
| 4 Repair | Entity, dated evidence, or product and fulfilment | Cross-functional | Each item has a deliverable and ship date |
| 5 Re-test | Same prompts, same market, two to four weeks apart | Growth | Two rounds agree before closing |
Four paths, not one:
- Factual error. Fix the entity first: name, category, founding year, origin, model numbers consistent everywhere. If the brand does not resolve to one node, nothing holds — start with whether your brand entity exists. Then a facts page with recall scope, affected SKUs, certificate numbers, and dates; identifiers beat “fully remediated.”
- Stale experience. Out-date it with dated evidence: release notes with firmware versions, an independent re-test, a support SLA with response times. Do not buy reviews or let an agency write “customers agree”; templated praise is recognizable and removable.
- Real defect. Say it plainly, publish a remediation timeline and compensation policy; concealment leaves the old thread as the only citable source.
- Malicious association or same-name confusion. Platform appeals and legal, not content.
Example: three months on one thread
Example — brand, product, and dates are invented; no real company is described.
A US smart-irrigation brand’s 2023 Reddit post — “app wiped my schedule; the lawn died” — was cited by two review sites. Firmware moved to local scheduling in 2024, yet answers still read “app is unreliable.”
Week one: a connection-change note on the site — models, firmware, change date, what failed — mirrored into structured data and the marketplace listing. Week three: units to two reviewers, no verdict conditions, only a request to state test date and firmware. Week five: a reply under the thread with versions and dates, not an argument. Week eight: the same prompt three times; the answer shifts to “early firmware had scheduling complaints; local scheduling added in 2024.” The negative did not disappear — it acquired a date, which is the realistic target.
If nothing moves, check cited URLs before adding budget: unchanged citations mean the material is not in the retrieval pool; changed ones with the same verdict mean it lacks detail. Sample as an incrementality experiment would.
Where this breaks
First, a real defect is not repaired by content. Recalls, safety failures, and false claims belong to compliance and product; washing an unfixed problem turns one thread into “brand response dodges the issue.”
Second, real reviews cannot be bought out. Deleting reviews and paying for coverage break platform rules and can breach EU consumer-protection law; aim for findable true evidence, not vanished bad evidence.
Third, answers fluctuate. Region, account, and hour change the output; one screenshot proves neither a fix nor a crisis’s end.
Fourth, small categories do not justify the machinery. A brand drawing a handful of mentions a year should spot-check quarterly, not run a stack.
Fifth, a live crisis runs in a different order. One agreed statement and a facts page precede ranking; a “brand refresh” release adds one more quotable sentence that says nothing.
This sits on the off-site authority line, closer to entities and reviews than to a publishing calendar.
This month, copy every negative sentence AI answers produced about you in the last 30 days, with its source URL and a classification. If one lands in “factual error,” fix the entity before drafting a statement, then re-run it in four weeks.