AI Reads What Buyers Say on Reddit and Trustpilot

Consider a composite scene. A DTC pet-gear brand is two weeks from shipping its first pallet to a German warehouse. The founder asks an assistant if the crate suits a dog with separation anxiety. The answer names two competitors and cites a Trustpilot page about a bent hinge and a community thread about a latch failing in cold weather. The brand’s own page — long FAQ, five-star average — is absent. Nobody wrote those sentences; nobody can delete them.

AI quotes the corpus around your brand, not your brand, so audit what that corpus already says before writing another page and spend only where a sentence can actually move.

Three corpora, three trust paths

Reviews, communities, and Q&A all read as “other people talking about us.” In an answer they do different jobs, so one response plan misfires.

Corpus	Who writes it	What the model takes	Common misread
Reviews	Media, test sites, affiliate roundups	Comparison: fit and price	A press release counts as coverage
Communities	Owners, hobbyists, forums	Lived detail: failures, support	Deleting a thread removes it
Q&A	Askers, answerers, platform Q&A	Narrow follow-ups	More answers always help

All three are judgments made in your absence. The difference is leverage: reviews can be pitched with a spec sheet, a community can be joined but not managed, Q&A can receive an official answer but never a buyer’s voice. Merge them into one PR bucket and you get a press push, an angry thread, and an answer box nobody has touched in years. This is the core of off-site authority — not content writing: you choose your pages, others choose the corpus.

Four dimensions, two of them slow

An audit is not a link list. It answers one question: can this corpus support “you should buy this”?

Dimension	Question	Failing signal
Volume	Independent sources per question	Your site plus one release
Quality	Fact or adjective?	“Great value,” no model or use case
Freshness	Last real discussion?	Old threads; two revisions since
Sentiment	Which complaint repeats?	One complaint everywhere

Only volume and freshness respond to outreach this quarter; quality and sentiment are set by product, packaging, and fulfillment. Skip the distinction and a GEO budget becomes a PR budget with no KPI.

Buying reviews to lift the average is the counterexample. A star rating is a platform field; the model extracts sentences. “Works great” adds nothing; “third charge, the contact pin failed, replacement took nine days” is lifted verbatim. Once rating and text disagree, the model reads contradiction.

Sort by leverage, not sentiment

A two-column sort, positive versus negative, produces a mood board. A three-column sort produces a work plan.

Bin	Examples	Do	Do not
Yours	Store reviews, official Q&A, sample units	Answer the question, publish specs	Incentivize ratings, refund for reviews
Shared	Media reviews, roundups, community threads	Verifiable material, one reply in your name	Post as a fake customer
Not yours	Competitor threads, old complaints, ranking	Record, watch, feed product	Demand removal, legal threats

Most teams spend ninety percent of their energy in the third column. The action there is reading, not fixing: packaging complaints belong to fulfillment, instructions to content, price to pricing. Content should log them, not own them, or the same complaint returns next quarter.

A one-week audit

Step	What	Who	Pass condition
1 Questions	20 from tickets, site search, competitor Q&A	Growth	8+ fit or comparison questions
2 Citations	Two or three assistants, one locale, one language	Content	Domain-by-question matrix
3 Sentences	Open each cited page, copy the sentence mentioning you	Content	A quote per row
4 Sort and act	Assign bins, schedule first-column actions	Growth and ops	Three actions, no new budget

Do not invent the questions. Eight “will this work in my apartment” items decide a recommendation more than twenty “what is it” items; pull them from the buyer question chain. Sample one market and one language, or regional differences read as corpus differences — see AI search’s four entry points.

First check the AI crawler map: a thread you can open may be closed to bots.

Example: one audit, two actions

Example: a hypothetical coffee-grinder brand preparing a UK launch runs 20 English questions. Its domain appears in 4 answers, a hobbyist community in 9, two review sites in 6. The recurring community line: “the burr wobbles after a few months.” The recurring review line: “the most complete accessory kit at this price.” Numbers are illustrative, not benchmarks; the sequence is the point.

Praise sits on review sites, risk in the community. First column: fix burr retention, publish a maintenance page with the revision date, send both review sites the accessory list. Second: one reply from the brand account — revision date, batch number, no argument. Third: keep the thread, comment once if cited again.

Eight weeks later that set returns “revised in the current batch” beside the thread, and three more question sets cite the brand. That proves the corpus moved, not revenue. Say so before an executive asks.

Where this breaks

The category has no third-party discussion. A new category or niche B2B line returns an empty sheet. Build owned evidence first; manufactured discussion has no history.

The corpus is unreadable. Closed platforms and crawl restrictions hide what buyers say. Sample manually and label the gap — “we could not see it” is not “nobody said it.”

There is a systemic trust problem. Mass complaints, regulator involvement, or chargebacks are not a GEO issue. Fix recalls, refunds, and compliance first; the next complaint buries corpus work.

The entity is not aligned. If one mention splits across three name variants, your volume number is fiction. Freeze name, category, and description first; the AI visibility baseline shows what to fix.

For this month, choose one action, run the 20 questions, open every third-party page, and copy the exact sentence into three columns. Ship the three first-column actions that need no new budget. Put the other columns in next month’s review, not this month’s OKR.