You've fixed everything — the reviews have changed, management has improved — but AI keeps presenting you to potential customers with the problems from two years ago. This isn't a technical lag that resolves on its own: that negative story is etched into the models' memory, and waiting changes nothing. Meanwhile, every customer who does their research before choosing you still reads the wrong version. There's a precise path to rewrite what AI says about you — and the results show up in months, not years.
A wellness hotel in South Tyrol went through 18 months of negative reviews after a change of management. Today it’s back to 4.8 stars, but AI keeps citing the old reviews in its answers. Here’s how we worked on the recovery.
I’m telling you about this case because it’s the most counterintuitive thing I’ve seen while working on visibility in AI answers: you can fix everything in the real world and in the Google world, but the entity in the Knowledge Graph stays dirty for months. ChatGPT and Perplexity keep pulling old signals, regurgitating them in their answers, and you can’t figure out why a potential customer is asking you to account for things that have been outdated for a year.
Entity recovery is a discipline of its own. It’s not PR, it’s not classic online reputation, it’s not SEO: it’s the work of rewriting the signals that AI models associate with your brand. Let me explain it through the real case I handled in South Tyrol, numbers included.
What I mean by “scars” in the entity
When an AI engine answers “best wellness hotel South Tyrol” it doesn’t read your site live. It pulls from an archive of representations — embeddings, attributes, relationships — built up over time. If for 18 months the signals tied to your entity were “chaotic management”, “rude staff”, “dirty spa”, those signals stay imprinted in the internal representation even after reality has changed.
In the world of research on retrieval and grounding systems, this phenomenon is called negative recency bias in corpora: models learn from the distribution of available documents, and if that distribution was dominated for a period by negative sentiment, mitigation requires a volumetric rebalancing, not just a chronological one. It’s not enough to wait for the old reviews to “age out”: you need new, abundant signals from high-trust sources.
I discussed this in the piece on how embeddings work in vector space: your entity is a point in a thousand-dimensional space, and every review, article, mention shifts that point. Eighteen months of pushes in the same direction leave a mark.
The case: wellness hotel, 80 rooms, province of Bolzano
Let me give you the context without revealing the brand. A 4-star superior hotel in a valley between Merano and Bolzano, 80 rooms, a 1,500 sqm spa, a clientele mostly German and from northern Italy. In 2024 the ownership changes, an unsuitable manager arrives, and in 18 months the Google rating drops from 4.6 to 3.2. Brutal reviews: cleanliness, food, staff. Ownership changes again in autumn 2025, an experienced director arrives, and in six months the rating climbs back to 4.8.
The problem: when a potential customer asks ChatGPT or Perplexity “wellness hotel South Tyrol for a week of relaxation”, the hotel either doesn’t show up, or shows up with caveats like “some recent guests report service problems”. Recent according to AI, two years old in reality.
Believing that responding to reviews is enough.
What we measured before the intervention
We ran a quantitative audit on a sample of 30 queries across three engines — ChatGPT, Perplexity, Gemini — simulating questions from a German and an Italian traveler. The pre-intervention results, in February 2026:
- Brand citations: 4 out of 90 (30 queries × 3 engines)
- Citations with neutral/positive sentiment: 1 out of 4
- Citations mentioning “management problems” or “mixed reviews”: 3 out of 4
- Direct competitors cited in the same set: 11 brands, with 34 citations overall
An indicative test, not a controlled study. But the pattern was unequivocal: the hotel was present but penalized in the models’ internal representation. Competitors in the same valley, with fewer rooms and a smaller spa, were getting 8x the citations.
Update all your structured data: Google Business Profile, the site’s Organization schema, the Wikidata entry if one exists.
The recovery plan in four directions
I’m not selling you a magic method. Entity recovery works when you push on several fronts at the same time, because AI models triangulate from different sources. If you clean up only Google but not Wikidata, or only the reviews but not the editorial content, the asymmetry remains.
A massive volume of positive content on third-party sources. We coordinated 14 editorial pieces over six months on high-end German and Italian travel publications — travel magazines, wellness guides, vertical industry portals. Not disguised advertorials: genuine reviews from hosted journalists, focused on the change of management and the new experience. This is the key point: you have to overwrite the old signal with a volume at least 3-5x higher than the normal pre-crisis cadence.
Updating all structured data. Google Business Profile updated with new photos, spa hours, a description explicitly mentioning the new management since 2025. The site’s Organization schema refreshed and tested with the Rich Results Test. The hotel’s Wikidata entry updated with the date of the management change as an explicit property. These signals are the backbone on which the models anchor the facts.
Cleaning up and contextualizing historical reviews. We publicly responded to the old negative reviews with a standard message that cites the change of management and invites a second, discounted stay. This doesn’t make the review disappear, but it adds a context signal that the models read together with the original text.
Weekly monitoring of AI answers. This is the part nobody does and the one that makes the difference. Every Monday morning, for 24 weeks, the same set of 30 queries across the three engines. Tracking how many times the brand shows up and with what sentiment.
The results at six months
Measurement from August 2026, same methodology, same query set:
- Brand citations: 27 out of 90 (up from 4)
- Citations with neutral/positive sentiment: 24 out of 27 (up from 1 out of 4)
- Citations referencing “past problems”: 2 out of 27, both contextualized with “under the new management”
- Average position in AI lists: from not present to top 5 in 18 out of 30 queries
Let me state the limit: a sample of 30 queries is not a rigorous longitudinal study, and AI engines vary their answers over time. But the pattern is clear: six months of coordinated work rebalanced the entity’s representation.
The mistakes I see most often in recoveries
I work with other hospitality businesses around Bolzano, Merano and the Val Pusteria, and these are the three mistakes I see recurring when someone tries the recovery on their own.
Believing that responding to reviews is enough. Replying to negative reviews helps, but it’s a weak signal. If the volume of new positive editorial content is zero, the models keep weighting the historical corpus.
Forgetting the sources by language. A hotel in South Tyrol that runs its recovery only in Italian ignores that 60% of the relevant mentions are in German. AI models answer in the language of the query, but they pull from corpora in all languages. If Spiegel, FAZ, German travel outlets never cite you, you’re left hobbling.
Confusing recovery with SEO. I’ve seen entire budgets spent on generic link building from low-quality sites. Entity recovery requires high-trust sources and semantic coherence, not link volume. Three reviews in specialized wellness magazines are worth more than thirty mentions on generalist travel blogs.
Not updating Wikidata and schema. If the public Knowledge Graph still reports attributes from the crisis period, you have an authoritative source contradicting the new narrative. As I told you in the piece on E-E-A-T for AI, coherence across sources is a trust multiplier.
What to do concretely if you’re in a similar situation
Operational steps, in order of impact on your visibility in AI answers.
- Measure the baseline: 20-30 real queries from your industry on ChatGPT, Perplexity, Gemini. Record citations and sentiment.
- Update all your structured data: Google Business Profile, the site’s Organization schema, the Wikidata entry if one exists.
- Plan a volume of editorial content on high-trust sources equal to 3-5x your normal cadence, for a minimum of six months.
- Publicly respond to historical negative reviews, contextualizing the change.
- Re-measure the same query set every 4-6 weeks. Without measurement you don’t know whether you’re pushing in the right direction.
These are entry-level steps: the real analysis of a post-crisis entity profile requires professional AI monitoring tools and a knowledge of your specific corpus of mentions. But starting here gives you 70% of the benefit.
The thread of visibility in AI answers
Entity recovery isn’t a luxury: it’s the necessary condition for all the rest of the GEO work to make sense. You can have the best content, the perfect inverted pyramid, your recognizable authors — but if your entity in the KG carries scars, AI models penalize you upstream of any evaluation of the content.
In this series on entities I’ve written before about how to build rich entity attributes and how to handle continuous entity monitoring. Recovery is the pathological case of that monitoring: when the monitoring flags an anomaly, recovery is the tool. In the next articles I’ll talk about how to handle the ordinary maintenance of the entity, how to react to weak signals before they become a crisis, and what to do when part of the brand’s story is objectively controversial and you can’t “clean it up”.
