Reputation management when the reviewer is a machine

For twenty years, online reputation management had a clear physics: search results are a ranked list, people rarely look past the first few, so the game was pushing good pages up and bad pages down. An entire industry grew around that displacement model.

AI answers do not work that way. When someone asks an engine "is [your brand] trustworthy," they do not get ten links to weigh. They get a verdict: two paragraphs, written with confidence, synthesized from everything the model absorbed in training plus whatever sources it just retrieved. The reviewer is a machine now, it has already read everything ever written about you, and it does not scroll. Managing your reputation with that reviewer requires understanding how it forms impressions in the first place.

Two memories: how a machine forms an impression

Every AI answer about your brand is shaped by two distinct layers, and they fail differently.

The long memory: training data. The model was trained on a huge snapshot of text: news coverage, review sites, forum threads, documentation, blog posts, complaint boards. Whatever the written record said about your brand across those years is compressed into the model's prior. This is why an engine can describe your company with no retrieval at all, and why that description sometimes sounds like your brand three years ago. It is also why a loud incident from 2022 can echo in answers long after the coverage died down: it is baked in until a newer model, trained on a newer record, replaces the old one.

The short memory: retrieval. Engines that search the live web (Perplexity always, ChatGPT and Gemini often, Google AI Overviews by construction) fetch a handful of current pages at answer time and lean on them heavily, usually with citations. This layer is faster-moving and far more tractable: change what the retrievable record says and the answers that depend on it can change within weeks.

The practical consequence: you cannot fix a trained-in narrative by editing your own website, and you cannot wait out a retrieval problem hoping the next model forgets it, because retrieval will keep re-finding the page. Diagnose which layer is producing the claim before choosing the fix. A claim that appears with a citation is retrieval; a claim that appears with no sources, phrased in generic past-tense terms, is usually the prior.

Audit: find out what the machine actually says

Reputation work that starts from vibes fails. Start from a systematic audit instead.

• Ask what buyers ask. Direct questions ("is [brand] good," "what are [brand]'s weaknesses," "[brand] pricing," "[brand] reviews"), comparative questions ("[brand] vs [rival]," "why do people switch away from [brand]"), and category questions where you should appear. Ten to twenty branded questions cover most of the risk surface.
• Ask every engine separately. ChatGPT, Claude, Gemini and Perplexity have different training snapshots and different retrieval behavior. It is normal for one engine to describe you accurately while another repeats a two-year-old complaint. Treat each engine as its own reviewer.
• Ask repeatedly. Answers are non-deterministic. A claim that appears once in ten runs is a low-probability tail; a claim that appears in seven of ten is your de facto public description. You cannot tell the difference from a single screenshot, which is why the audit needs repeated sampling, with sample sizes attached to anything you report internally.
• Capture the citations. For every retrieved claim, record the source. This turns "the AI says our support is bad" into "this specific review page from 2023 is cited in 60% of support-related answers," which is a task, not a grievance.

Then triage what you find into three buckets: factual errors (wrong pricing, dead products, departed executives), stale narratives (true once, not anymore), and hostile framings (a real criticism given more weight than the record supports). All three respond to the same playbook at different intensities.

The counter-playbook: three moves, in order

Move 1: correct the source

Because retrieval-driven claims cite their inputs, the fastest fix is upstream. Trace the claim to its cited pages and work through them:

• Pages you control: fix them today. Stale pricing pages, outdated docs and abandoned product pages are self-inflicted reputation wounds, and they are the ones engines cite most confidently because they come from your own domain.
• Pages where you have standing: review platforms, comparison sites and directories usually have vendor-response or correction mechanisms. A factual correction with evidence frequently works, and even a public vendor response changes what a retrieving engine reads on that page.
• Pages you can only influence: journalists update articles when the facts changed and someone politely shows them. Old coverage of a resolved incident often accepts an update note. This is classic PR work with a new payoff: the article is no longer just read by people, it is read into answers.

Move 2: publish the counter-evidence

Engines can only retrieve what exists. If the only detailed page about your 2022 outage is the post-mortem coverage, that is what gets retrieved forever. Publish the page you want retrieved instead: direct, specific, well-structured, addressing the claim head-on. "What happened in 2022 and what we changed" outperforms silence. So does a straight pricing page against "their pricing is opaque," and a documented migration guide against "switching away is painful."

Two rules make counter-evidence work. Address the claim explicitly, in the language people use when they ask about it, because a page that dodges the topic will not be retrieved for it. And make it verifiable: dates, numbers, named changes. Engines synthesize across sources, and a specific rebuttal survives synthesis next to a vague criticism far better than corporate boilerplate does. Structure matters too; the mechanics are in our guide to what makes a page AI-readable.

Move 3: earn third-party validation

Here is the uncomfortable part: your own site is the least trusted voice on the question of whether you are good. In our citation analysis, engines overwhelmingly cited independent sources over brand-owned pages. A synthesized answer weighing "the vendor says they fixed it" against "three independent sources say it is still broken" will side with the three sources, and it should.

So the durable fix is a corrected independent record: fresh reviews from current customers on the platforms engines cite for your category, analyst and press coverage of what actually changed, credible third parties telling the current story. This is the slowest move and the one that actually rewrites both memories, because the independent record is what future models train on.

Why SEO-era suppression fails

The classic ORM toolkit was built to exploit a ranked list: flood the index with positive pages, push the bad story to page two, and rely on nobody looking there. Against a generative engine, every mechanism in that toolkit misses.

• There is no page two. An answer is a synthesis, not a list. The negative story does not need to rank first to shape the verdict; it needs only to exist in training data or be retrievable at all.
• Volume does not dilute. Ten thin positive posts do not outweigh one substantive negative source in a synthesis. Models weigh specificity and corroboration, not page count.
• The long memory does not forget on command. Once a narrative is trained in, no amount of burying changes the prior. Only a corrected record, absorbed by the next training run, does.
• Astroturfing is a compounding risk. Fake reviews and sock-puppet posts become part of the permanent record. If they are ever exposed, the exposure becomes retrievable and trainable too, attached to your name.

The honest summary: against a machine reviewer, the only reliable strategy is making the true, current record the best-documented one. That is less like suppression and more like litigation: you win on evidence.

Realistic timelines

Anyone quoting a guaranteed fix-by date for an AI reputation problem is guessing, because the timelines belong to systems nobody outside the labs controls. What experience supports:

• Your own pages: corrections can enter retrieval within days to weeks of a re-crawl.
• Retrieval-driven claims generally: weeks is the honest unit, as engines re-crawl the corrected sources and re-cite. Watch when the updated page first appears in citations; that is your leading indicator.
• Trained-in impressions: model release cycles, typically months. The correction has to exist across the record long enough to be in the next snapshot. Progress is stepwise: flat, then a new model ships, then the framing changes.

Because the timelines are uncertain, measure instead of promising. Baseline the offending claims before you act (how often each appears, per engine, with sample sizes), do the work, and read the trend over elapsed time with confidence intervals rather than declaring victory off one clean answer. In llemmy this is what the Sentiment Tracker and a Campaign are for: the branded-question framing tracked per engine, the fix dated on a timeline, and the before-and-after read against a 95% interval. One clean answer is a coin flip; a framing shift that holds across weeks of sampled answers is a fixed reputation.

FAQ

How do AI engines form an impression of a brand?

Through two layers: a long memory absorbed at training time from years of articles, reviews and forum threads, and a short memory retrieved at answer time from live pages. Answers blend both, which is why fixing every live page can still leave an old narrative echoing from training data until a newer model ships.

How do you audit what AI engines say about your brand?

Ask each engine the questions buyers ask (direct, comparative, category), repeatedly, and record answers with their cited sources. Audit per engine, because impressions differ, and trace every wrong claim to the page feeding it so the fix is a task with a URL, not a grievance.

How do you counter a negative narrative in AI answers?

Correct the source, publish the counter-evidence, earn third-party validation, in that order. Fix pages you control, use correction mechanisms where you have standing, publish direct verifiable pages addressing the claim, and rebuild the independent record engines actually trust. Suppression does not work: an answer has no page two.

How long does it take to change what AI says about your brand?

Retrieval-driven claims can move in weeks as engines re-crawl corrected pages; trained-in impressions move on model release cycles, typically months. No one can promise dates, so baseline first and measure the shift over elapsed time with sample sizes and confidence intervals.

By the llemmy team, July 2026. Related reading: Brand monitoring has a new surface: AI answers, What AI engines actually cite, and How llemmy measures AI visibility (and what we don't claim).