Why Your Five-Star Reviews Are Invisible to AI

Two years ago, 25% of consumers used AI tools instead of search engines to find products and services. Today, that number is 58%.

Despite this jump, most businesses still optimise their content strategies in traditional ways. As a result, companies that dominated the first page of Google were sometimes absent from AI answers entirely.

At Alps2Alps, I have spent months testing to understand what was happening. With the team, we rebuilt our strategy, and our systematic GEO work put us at the top among competitors in AI-searched models (based on the Basebright service).

The short thought is that AI doesn’t rank websites. It compiles responses from diverse sources on the internet, with criteria for evidence that differ fundamentally from SEO standards.

From search rankings to AI citations

Researchers at Princeton and Georgia Tech coined the term Generative Engine Optimisation (GEO) in a 2024 study published at KDD, one of the top data science conferences. They tested how different content strategies affected visibility in AI-generated responses and found that the right optimisations could boost citation rates by up to 40%. They also found that traditional SEO tactics like keyword stuffing actually hurt performance in generative search.

The business case is straightforward. Gartner predicts traditional search volume will drop 25% by 2026 as users migrate to AI assistants. Capgemini’s 2025 consumer research, covering 12,000 people across three continents, found that 58% have already replaced search engines with generative AI for product discovery. According to Statista, about 40% of travellers now use AI tools to plan trips.

These aren’t projections about some distant future. This is already how a growing share of your customers find you or don’t. After a year of testing and rebuilding around this reality, patterns have emerged. After countless tests and rebuilding the content strategy to the new reality, here is what works and doesn’t.

Perfect reviews are the new red flag

The Princeton study found something counterintuitive. Adding quotations and concrete statistics to content improved AI visibility far more than polishing the language or stuffing it with keywords. The implication for businesses that rely on customer reviews is significant. That’s not to say star ratings don’t matter, because AI reads the distribution of scores on trusted platforms like Trustpilot and Google Maps, and a strong aggregate signal counts. But it only gets you so far. When a model synthesises an answer about which transfer to book, it needs something to extract. “Great service!” is not that. This makes sense if you think about how these systems work. When ChatGPT answers a question about transfers for travellers, it doesn’t just count stars. It synthesises information from review platforms, forums, social media, and editorial content. Detailed accounts give the model something concrete to work with. This makes sense if you think about how these systems work. When ChatGPT answers a question about transfers for travellers, it doesn’t just count stars. It synthesises information from review platforms, forums, social media, and editorial content. Detailed accounts give the model something concrete to work with.

The uncomfortable part is that some conflict in your review profile actually helps. For example, travelling involves weather disruptions or flight delays. When those problems appear in public reviews and a company responds with specifics rather than a template “we value your feedback,” it signals authenticity. A profile with nothing but praise looks manufactured to consumers and to AI systems that weigh source credibility.

The practical takeaway is simple but hard to put into practice. Stop filtering who you invite to leave reviews. Ask every customer, including those who had a rough experience. Respond to criticism publicly with concrete details about what happened and what changed. The messiness is the point, because it’s what separates earned reputation from synthesised.

Write for extraction, not impression

Most marketing copy is written to persuade humans browsing a website. GEO requires writing for extraction. Your content needs to be structured so that an AI can pull out specific facts and cite them in a combined answer.

Language models pull from multiple sources when generating an answer and favour independent platforms. A mention on TripAdvisor, a Reddit thread where a real user described their experience, a YouTube video from an influencer who actually took the service carries more citation weight than the same information on your corporate blog. User-generated content on third-party platforms indexes well in generative search because it’s detailed, conversational, and comes from an independent source.

The same logic applies to earned media. If an industry publication quotes you in an article about your sector, that quote becomes citation-ready material for AI. It sits on an authoritative domain, it’s attributed to a named expert, and it addresses a specific topic. One good quote can surface in dozens of AI-generated answers.

On your own website, the changes are more technical but equally important. Structured data markup helps AI parsers understand what your business does, where it operates, and what services it offers. FAQ sections with direct, specific answers perform well because they match the question-answer format that language models work in. And the language itself matters: “We are the leading service” is noise. “We operate 12 services across 5 countries with an average response time of 6 minutes” is data that AI will use.

Language coverage is an important factor that is often overlooked. AI responds in the language of the query. If your website exists only in English but half your customers search in French, German, or Italian, you’re invisible to those queries. For any business operating across multiple markets, multilingual content is a GEO requirement.

Your support tickets are a content goldmine

Most companies that use AI have someone on the team with a ChatGPT tab open for basic tasks. That’s a starting point, not a strategy. The real competitive advantage in GEO comes from using your own operational data to guide what you create and optimise.

B2C services can handle thousands of customer queries every month. In transfers, for example, about 45% of these are repetitive. When we started routing them through an AI assistant, the immediate obvious benefit was faster response times and freed-up support staff for difficult cases. But another important benefit was the data. We now have a clear map of what customers actually ask, in their own words, at every stage of the journey.

Those questions are the same ones people will ask ChatGPT. If you have a detailed, specific page answering exactly that, you become the source AI cites. If you don’t, someone else’s Reddit comment becomes the answer instead.

The same principle applies to competitive monitoring. Which sources get cited? Where are the gaps? When a new competitor starts appearing in AI answers, what content did they create that earned the citation? And this should be a weekly practice for businesses.

GEO is a continuous process, not a project with a launch date. Generative models update constantly. The sources they prioritise can shift. What earned you a citation three months ago might not work today. The companies that build a feedback loop will compound their advantage over time. Everyone else will keep wondering why their Google rankings don’t translate into AI visibility.

A year ago, I didn’t think much about how ChatGPT described our industry. Now it’s one of the first things I check every week and answers change. New sources appear, old ones drop out. It’s a living system, and the only way to stay visible in it is to keep feeding it something real.