You have twenty positive reviews and three negative ones on an authoritative outlet — and you think the ratio is in your favor. For AI it doesn't work that way: three negative comments on a high-weight source count for more than twenty positive ones on generic directories, because the model doesn't average them out, it weighs the authority of the source. And those negative mentions already acquired by the AI can't be erased with legal requests or technical redirects: they're there to stay. The only lever that works is building a higher volume of positive mentions on equally authoritative sources — and there's a fifteen-minute test that tells you where you stand today.
You have 20 positive reviews and 3 negative ones. For Google that’s fine. For AI, if those 3 are cited in an important article, they weigh three times as much. Here’s how to dilute.
The math of reputation is no longer linear now that ChatGPT, Perplexity, Claude and Gemini are also reading the web. The AI engine doesn’t take the arithmetic average of reviews the way Google does: it weighs every mention by the authority of the source, the semantic context in which it appears, and its co-occurrence with other entities. Three negative mentions on an authoritative local outlet weigh more than twenty positive reviews on generic directories.
In my articles in this series I’ve already explained how the citation economy works for AI: it’s not just the link that counts, it’s the mention in context. Today I’ll explain the part everyone avoids — what to do when a negative mention is already out there, can’t be removed, and is ruining how the AI describes your brand.
Why you can’t simply delete it
The negative mention lives in three places at once: on the original source, in the search engines’ cache, and — worst of all — in the training corpus of the AI models. When Anthropic, OpenAI or Google crawled the web to train their models, they took everything. Including that three-star review on a tourism aggregator that complained about the breakfast.
Even if you remove the source page today, that mention is already baked into the model’s parameter weights. It can’t be extracted. It can’t be undone with a legal request. It can’t be erased with a 301 redirect.
The only lever you have left is the volume and authority of your future positive mentions. The model, at its next retraining or its next real-time retrieval via search integration, updates the aggregate sentiment based on what it finds. If it finds ten times more positive signals than negative ones, on sources of equal or higher authority, the balance shifts.
This isn’t a classic reputation-management trick. It’s recognizing how the architecture of the system reading you works when a user types “farm stays rural villages central Sicily” into ChatGPT.
How the AI model reads aggregate sentiment
In previous articles on how AI engines think I explained that AI doesn’t read texts, it reads tokens in context. When a user asks for an opinion about your brand, the model recalls all the passages in the corpus where your name appears, weighs them, and synthesizes an answer.
The weight of each passage depends on three variables I’ve seen emerge clearly while working with brands in reputational crisis:
- Source authority: a mention on an industry outlet weighs 5-10 times more than a mention on a generic forum. I wrote about it in backlinks as citation proxy.
- Co-occurrence with positive entities: if your brand appears next to “award”, “conservation restoration”, “recognition”, the model learns a positive association.
- Recent temporal frequency: mentions from the last 12 months weigh more than old mentions. The model prefers fresh information.
It follows that dilution isn’t sweeping things under the rug. It’s saturating the semantic field with higher-quality positive signals, so that when the AI builds the aggregate sentiment, the negative weight becomes a negligible fraction of the total.
Responding publicly to the negative review in a combative tone.
The test you can run in 15 minutes
Before planning any dilution campaign, you need to know what the AI thinks of your brand today. It’s not a technical audit, it’s a conversation.
Open Perplexity and run three queries about your brand:
- “What do people say about [brand name]?”
- “[brand name] opinions and reviews”
- “Experiences with [brand name]”
Then repeat the same three queries on ChatGPT with search enabled and on Gemini. For each answer, note: which sources it cites, in what tone, whether it mentions negative aspects and where it got them from.
If the same negative aspect comes up on 2 out of 3 engines, you have an urgent dilution problem. If it appears on only one, you’re in a manageable phase. If none of them cite it but you know the negative mention exists, you’re in a window of opportunity — you need to move before it enters retrieval.
To map the overall volume of your mentions you can also use Google Trends to understand when people are talking about you and Google Search Console to see which pages receive queries that mention your brand negatively. It’s an entry-level check: real analysis requires professional brand-monitoring tools.
For every Tier 1 mention, plan 8-10 positive mentions on Tier 1 or Tier 2 sources over the following 6-9 months.
The case of the farm stay in the village
Let me tell you a concrete story, because theory without a case gets forgotten.
An operator in the province of Enna had restored an abandoned rural village and turned it into a scattered farm stay — houses spread across the historic core, a restaurant in the old stable, a wine cellar beneath the former oil mill. A serious project, years of investment. Then, in the summer two years ago, a difficult season: three strongly negative reviews on a reasonably authoritative travel outlet, amplified by an article in an online magazine specialized in slow tourism that picked up those complaints.
When the operator contacted me, he had run the test on Perplexity: to the query “restored village farm stays central Sicily”, the model listed five options. His village appeared in fourth place, with a note summarizing “some issues with service and cleanliness reported by recent reviews”. Translation: the AI had learned the negative frame and was repeating it.
The intervention was structured dilution over 9 months. The goal: generate at least 8 positive mentions on sources of equal or higher authority for each individual negative mention detected. These aren’t fake reviews, they’re real PR activities:
- Three editorial features on national experiential-tourism outlets, obtained by offering two journalists a reportage stay.
- A participation in a RAI documentary on the restoration of Sicilian villages.
- Four pieces in architecture and restoration magazines, focused on the conservation work.
- Two speaking appearances at sustainable-tourism conferences, with proceedings published online.
- An improved entry on Wikidata with links to the authoritative sources just obtained.
After 9 months we re-ran the same queries on Perplexity, ChatGPT and Gemini. The brand had risen to first place on two engines, and second on the third. The negative note had disappeared from the summary: it had been replaced by phrases like “celebrated for the conservation restoration of the village”. The three original negative reviews still exist, they’re right where they were. But they now weigh a fraction of the total.
This is a single case study, one operator, one specific sector: an indicative pattern, not a statistical study. But the logic of the mechanism is documentable and reproducible in other sectors.
The mistakes I notice most often
After 18 months of work on brands with negative-mention problems, the patterns that always come back are four.
Trying to get the original source deleted. It rarely works, it leaves legal traces worse than the initial problem, and above all it doesn’t remove the mention from the already-trained corpus. A waste of energy.
Responding publicly to the negative review in a combative tone. It generates additional textual context around the negative mention, reinforcing the association in the model. Your brand becomes co-occurrent with “conflict”, “defense”, “controversy”. The AI learns the frame worse than before.
Generating low-authority positive mentions in volume. Twenty listings on bottom-tier directories don’t weigh as much as a single article in a Tier 1 outlet in the sector. The AI doesn’t count, it weighs. One strong signal beats ten weak ones, as I explained in implicit reference weight.
Moving only when the crisis is acute. Dilution takes 6-12 months to affect the aggregate sentiment of the models, because it has to pass through the retraining cycles or the update cycles of the search-integration indexes. Starting late means living with the negative perception for quarters.
What to do concretely
If you’ve detected a relevant negative mention, here’s the operational sequence:
- Classify the authority of the negative source on a three-tier scale: Tier 1 (national/industry outlet), Tier 2 (regional/vertical), Tier 3 (aggregator/forum).
- For every Tier 1 mention, plan 8-10 positive mentions on Tier 1 or Tier 2 sources over the following 6-9 months.
- For every Tier 2 mention, 4-5 positive Tier 1 or Tier 2 mentions in 3-6 months.
- For Tier 3 mentions, 2-3 positive Tier 2 mentions are enough — the source’s authority is already low.
- Re-run the quarterly test on the three Perplexity queries to measure the shift in sentiment.
Always benchmark against the 3-5 competitors the AI cites in your sector when you run the query “best [category] [area]”: you’ll see which sources the AI considers credible and you’ll be able to target them.
Where the discussion continues
Dilution is a long-term lever in the work of visibility in AI answers. It’s not a magic factor, it’s not enough on its own, and it doesn’t replace an overall PR strategy. It only works when integrated with quality editorial production and a presence on sources the models already consider authoritative.
In the next articles in this series I’ll cover two complementary mechanisms: citation attribution accuracy, that is, how to make sure the AI correctly attributes sources when it cites you, and social mention aggregation, which explains how social signals flow into the aggregate sentiment the model uses to describe you.
