Why AI Chatbots Recognise Startups But Rarely Recommend Them

A new study shows AI chatbots almost always recognise a startup when asked about it by name, but rarely recommend that same startup when asked an open question like "what's the best new AI tool". For ChatGPT, the gap between recognition and recommendation is roughly 30-to-1. That means being technically visible to AI is a very different problem from being recommended by it, and most GEO tactics only fix the first one.

What changed?

Research firm ReachLLM this week highlighted a January 2026 academic study testing how large language models handle discovery-style queries versus direct, name-based queries. The paper, The Discovery Gap: How Product Hunt Startups Vanish in LLM Organic Discovery Queries, examined 112 Product Hunt startups across 2,240 prompts, according to Indiana Headlines.

The results, drawn from the original arXiv paper by researcher Amit Prakash Sharma, are stark. When users asked about products by name, both models recognised them almost perfectly: 99.4% for ChatGPT and 94.3% for Perplexity.

That recognition evaporates the moment the question becomes open-ended. When users asked discovery-style questions such as "What are the best AI tools launched this year?", success rates collapsed to 3.32% for ChatGPT and 8.29% for Perplexity, a gap of roughly 30-to-1 for ChatGPT specifically.

Perhaps the most uncomfortable finding for the GEO industry: applying generative engine optimisation techniques to a company's own website didn't move the needle. The paper found GEO optimisation showed no correlation with discovery success on its own, suggesting it works as a multiplier only once other signals, such as SEO authority and community presence, are already established.

The study also draws out an architectural distinction worth understanding. ChatGPT (gpt-4o-mini) is a knowledge-cutoff model: it knows what it knew at training time and nothing more. Perplexity (sonar with web search) is search-augmented: it queries the live web and incorporates current information into its answers. The gap between the two models' discovery rates suggests that whether an AI system searches the web live, or simply recalls its training data, has a huge effect on whether new brands get a look-in at all.

Why does this matter for your business?

This research answers a question a lot of founders and marketers have been asking all year: if my site is technically optimised for AI crawlers, why doesn't ChatGPT ever mention us unprompted? The blunt answer is that structural factors, not on-page optimisation, decide whether a new or lesser-known brand gets surfaced in a broad recommendation.

The researcher's own conclusion is worth taking seriously. The architecture divide between knowledge-cutoff and web-search models emerged as the central finding with practical implications. For startup founders, the strategic implication is counterintuitive: don't optimise for AI discovery directly. Build the traditional foundations of SEO authority, community presence and platform success, and AI visibility tends to follow. As the paper puts it, the cart cannot pull the horse.

That's a meaningful correction to a narrative that has built up around GEO over the past year, that publishing "AI-friendly" content, adding schema markup, or chunking pages into extractable blocks will get a brand recommended by chatbots. This study suggests those tactics may help once you're already established, but they can't manufacture discovery from nothing.

The wider context supports this. Other 2026 research covered by Adapt Worldwide has already shown big shifts in where AI referral traffic lands, and separate industry analysis found that Google's AI Overviews, now appearing in roughly one in five Google searches, have reduced organic click-through rates by an estimated 61% when they appear, according to Indiana Headlines. Put together, the picture is one where fewer clicks are available overall, and the ones that remain go disproportionately to brands with existing authority, not to whoever has the best-optimised FAQ block.

What should you do now?

First, stop treating "AI visibility" as a single problem with a single fix. Being recognised when someone names your brand directly is a different challenge to being recommended when someone asks an open question in your category. The study shows these move independently, and most GEO tactics only affect the first.

Second, invest in the slower-moving signals the research points to: genuine SEO authority, real community presence such as Reddit threads, review sites and forums where your product is discussed unprompted, and a track record that predates the query. These are the inputs that fed visibility in the study, not schema markup or content freshness alone.

Third, treat model architecture as part of your strategy. A search-augmented assistant like Perplexity can, in principle, pick up new information about your brand as soon as it's published online. A knowledge-cutoff model like the version of ChatGPT tested here cannot, until it's retrained or unless it's paired with a live search layer. Knowing which type of system your prospective customers are using changes what "getting cited" even means for you.

Finally, measure the right thing. If you only track whether AI tools recognise your brand name, you'll miss the far more important question of whether they recommend you when a customer hasn't already heard of you. That's exactly the kind of gap worth checking for your own brand, and running a free AI visibility audit with Sited is a straightforward way to see where you actually stand across both direct and discovery-style queries.

Frequently asked questions

What is the "discovery gap" in AI search?

It's the difference between an AI system recognising a brand when asked about it by name, and the same AI recommending that brand unprompted in response to a broader question. The new study found this gap can be as large as 30-to-1 for ChatGPT.

Does GEO (generative engine optimisation) still work?

The study found that GEO optimisation, the set of techniques proposed for improving AI visibility, showed no correlation with discovery success on its own. It appears to amplify existing authority rather than create visibility from scratch, so it isn't wasted effort, but it isn't a standalone fix either.

Why does Perplexity recommend new products more often than ChatGPT?

Perplexity uses live web search to inform its answers, while the version of ChatGPT tested relies on training data with a fixed knowledge cutoff. Both recognised named products almost perfectly, at 99.4% for ChatGPT and 94.3% for Perplexity, but the search-augmented model was notably better at surfacing newer brands in open discovery queries.

What should a new startup actually do to improve its chances?

The research suggests focusing on traditional SEO authority, genuine community presence on platforms like Reddit and review sites, and building a track record over time, since these factors predicted visibility rather than AI-specific content tricks.

Is this study representative of all industries, not just startups?

The dataset was limited to 112 Product Hunt startups across 2,240 prompts, so results may vary by sector and by which AI platforms customers actually use. It's a strong signal rather than a universal rule, so brands should verify their own position with direct testing.