AI Reputation

Lesson 01

Where AI gets its opinion

An AI answer about a brand is a mirror of its footprint. Five source families feed it: your own site, reviews, news, knowledge bases such as encyclopedias, and forums and social posts. When they agree and are current the reflection is sharp; when they clash or go stale it warps. Every claim leaves a trail back to its sources, so trace a claim before you react. Different questions lean on different sources: feelings ("is it quiet?") lean on reviews and forums, facts ("what time is check-in?", "who owns it?") on your site, news and knowledge bases. The source mix tells you where to work. Worked example: two years after Quillmere Harbour’s renovation ended, assistants still warn about building noise, fed by 2023 reviews and a travel-forum thread, while the hotel’s own site never said the work was done.

Lesson 02

Facts and feelings

Answers mix two kinds of claim. Fact claims can be checked: a check-in time, an owner, a price, a date. Each is right or wrong and traces to one or two sources. Feelings are summed up from many voices, mostly reviews and forums: "tired", "noisy", "welcoming". The two are fixed differently. A wrong fact is corrected at its source. A bad feeling cannot be argued away; it is earned back with better experiences, answered complaints and fresh evidence. The claim ledger records every claim, its type (fact or feeling), its source family and its status, so the team always knows which kind of fix it is working on.

Lesson 03

The reputation audit

An audit replaces screenshots with a method. Build a reputation prompt set: the questions buyers and journalists ask about the brand by name ("is Quillmere good?", "Quillmere reviews", "Quillmere or Caskwell?", "who owns Quillmere?", "any problems at Quillmere?"). Ask each prompt in several assistants, more than once, in clean sessions, and date every run. Grade every claim right, stale, wrong or missing. Then rank: severity is reach (how many prompts and assistants carry the claim) times harm (what it costs a buyer or the brand). The audit sheet and the top-five list are what the team fixes first, and the baseline the next audit is compared with.

Lesson 04

Fix it at the source

A wrong claim is fixed where it comes from, so the correction ladder starts at home. First your own page: correct it and date the change. Then the profiles and directories that repeat the fact: business profiles, maps, booking and listing sites. Then the publisher: ask for a dated update note on the original article. Then the thread: reply with the current fact, saying who you are. Only then the assistants’ feedback buttons: the thumbs-down, the flag, the report form. As of October 2026 those are signals, not a correction desk: vendors say they use feedback to improve their systems, and none promises to correct an answer about a business on request. Personal-data removal requests exist for people under privacy law, not for brand facts.

Lesson 05

Your official facts hub

Assistants are confident when the facts agree everywhere, so give them one place to agree with. The facts hub is a plain-text page on your own site that states every core fact with an as-of date: legal name and brand name, founding year, owner, locations and addresses, hours, prices, policies, and what changed when. Ask your web team to label the page as an Organization (a machine-readable label called structured data) with links to your official profiles, link the hub from every profile, and keep every profile in agreement with it. Facts locked in PDFs, images and press kits are weak copies that drift out of date. Files such as llms.txt and brand.json can point to the hub, but as of October 2026 they are proposals, not standards.

Lesson 06

Entities and knowledge panels

To an assistant, your brand is an entity: one distinct thing with a stable name and identifiers. A same-name company, a rival next door or an old trading name can blur it, and a blurred entity gets merged facts and the wrong photo. Tell yourself apart with the legal name, the address, the founding year, registration numbers and links to your official profiles, on the hub and on every profile. A knowledge panel is built by the search engine from sources across the web: a verified representative can claim it and suggest changes, which are reviewed against public information (a description taken from another site is changed at that site). Wikipedia and Wikidata follow their own rules: an article needs significant coverage in independent sources (press releases do not count); people with a conflict of interest are strongly discouraged from editing the article and should disclose and request changes on its talk page; anyone paid to edit must disclose it under the Wikimedia Foundation’s terms of use.

Lesson 07

Honest review programs

Reviews are the loudest voice in tone, and the most regulated. An honest program asks every customer after the experience (not only the happy ones), makes it easy, replies to every review with the facts and the fix, never ties a reward to what a review says, discloses any reward, never writes or commissions reviews, and never threatens reviewers. The rules, by jurisdiction and date: in the United States the FTC’s rule on consumer reviews and testimonials (in force since 21 October 2024) and its Endorsement Guides (revised July 2023); in the United Kingdom the Digital Markets, Competition and Consumers Act, which since 6 April 2025 bans fake reviews, reviews that hide a reward, and publishing reviews in a misleading way; in the European Union the Unfair Commercial Practices Directive as amended, applied since 28 May 2022 through each member state’s own law, which bans false reviews and misrepresenting reviews. Platforms add their own, often stricter: as of October 2026 Google allows asking but bans rewards and selective asking, and Yelp says businesses should never ask. Rules change: check your local rules and each platform’s policy before you launch. This is not legal advice.

Lesson 08

Earned media and experts

Assistants lean on independent, named, dated coverage when they decide whether to trust a brand: news, trade press, expert commentary. The brand’s own voice (press releases, site claims, the chief executive’s posts) is useful for facts but weak as proof that others agree; Wikipedia does not count it towards notability at all. Earned media is won with a story a journalist can check: new, specific, timely, with data and a named expert. Keep an experts page on your site with names, roles and credentials, so journalists and crawlers find them. Several independent sources saying the same thing is corroboration, and corroboration is what makes a mirror sharp. Paid coverage must be labelled as paid, and it is not independent.

Lesson 09

Communities, with disclosure

Forums, question-and-answer sites and social groups are where feelings and recommendations form, and industry studies in 2025 and 2026 put community sites among the domains AI answers cite most. The rule is to join as yourself. Say who you are and what your interest is in the first line. Read each community’s rules first: many ban or ration self-promotion, and those rules come before your plan. Answer questions, add facts, reply to threads about your brand. Never astroturf: no fake accounts, no staff posing as customers, no paid posts without disclosure. In the United States the FTC’s Endorsement Guides expect paid and employee endorsements to be disclosed; EU and UK consumer law ban a business posing as a consumer and paid promotion that is not made clear. Check your local rules.

Lesson 10

When old crises linger

A resolved incident keeps reflecting while its coverage is still the freshest, most-linked thing about you, and a correction never fully erases a first impression (researchers call this the continued influence effect). The fix is not deletion. Keep the record and add the resolution: a dated resolution on the facts hub, update notes on the original coverage, fresh and specific coverage of what is true now, and a direct answer in your own voice to the question people ask. Write statements as a truth sandwich: the fact, the myth once, the fact again with evidence. Match the response to the responsibility: deny only what is false; when you were responsible, acknowledge it, show the fix with dates and rebuild.

Lesson 11

Fakes, errors and rivals

Three different problems need three different answers. Fake reviews and endorsements, for you or against you, show patterns: a burst in one night, near-identical wording, brand-new accounts. Document the pattern and report it to the platform with evidence; never answer fakes with fakes, and remember that paying for negative reviews of a rival is banned too. AI errors come in two kinds. A claim that appears in no source is a hallucination (the model guessed): send feedback and make the true fact easy to find. A claim that appears in a source is a source error: fix the source. Rivals are handled with honest comparisons: specific, checkable facts on your own pages, no running the rival down, no posing as someone else. Whether an AI provider answers in law for a false statement in an answer is being tested in the courts (a German court said yes in a first, not yet final, ruling in May 2026); that is a question for a lawyer, and the last rung of the ladder.

Lesson 12

Watch, escalate, respond

Reputation work becomes a routine. Watch: a monthly mirror check runs the reputation prompt set in several assistants, with repeats, in clean sessions, against last month’s baseline. Set triggers with thresholds: a wrong core fact in two assistants, a drop in tone on a key prompt, a new source in the mix. Escalate by rule, up a ladder with an owner and a time limit on every rung: watch, verify (reproduce it, grade it, find the source), fix the source, send feedback, make a public statement, go to a lawyer. Respond with facts, in the truth-sandwich form. Report monthly: what assistants say now, what changed, what was fixed and what is open, with sources. The course ends where it began, in a hall of mirrors that is finally sharp.