TLDR
- AI models, not search rankings, are becoming the new gatekeepers. A brand can rank highly on Google yet still be absent from chatbot responses if it lacks sufficient representation in training data, according to geoSurge CEO Francisco Vigo.
- Text visibility matters more than ever. Brands that reduced text publishing in favor of video and multimedia risk being “forgotten” by models trained primarily on text.
- SEO is evolving, not dying. As AI-driven discovery grows, brands must rethink analytics, avoid mass synthetic content, and focus on shaping how models “remember” them.
As consumers increasingly turn to AI chatbots rather than traditional search engines to discover products and services, brands face a growing risk of disappearing from AI-generated answers – even when they still rank well on Google.
That is the problem geoSurge is trying to solve, according to Francisco Vigo, CEO and co-founder of the U.K.-based startup, who says large language models can quietly ‘forget’ brands when training data changes.
GeoSurge was founded after Vigo became frustrated with ad-heavy search results while shopping online. “I wanted to buy some gym shoes. I went into Google, saw the first seven links were ads, and I was a little bit (ticked) off,” he said in an interview with The AI Innovator. “I trusted more what GPT was telling me than what search engines were surfacing.”
Vigo, a former lead data scientist at a $2 billion fintech company, said that moment made him realize brands would need new tools to understand how AI systems surface information. “That was kind of my ‘aha’ moment where I’ve realized that brands would be desperate to be positioned inside of how these models are surfacing information,” he said.
Why chatbots ‘forget’ brands
Unlike traditional search engines, chatbots rely on large language models trained on massive text datasets. Vigo said that means brand visibility is tied not just to SEO, but to how frequently a brand appears in training data. “Without the data, I won’t learn anything,” he said.
GeoSurge tracks the same prompts multiple times a day across 50 locations to observe differences in AI model results when new versions are released. That approach revealed what Vigo describes as a “massive” risk for brands.
For example, “on GPT-4, Chanel was consistently ranking among the top 10” most visible brands, he said. “When GPT-5 dropped and we started fetching data for GPT-5, Chanel completely disappeared from the memory of the model.”
What happened was that when GPT-5 was trained, “the footprint of Chanel in the training data set got diluted, so they were less visible. Therefore, when the model was doing its next word prediction, Chanel was not the most likely outcome” and thus was mentioned less, Vigo said.
Chanel didn’t do anything wrong. Instead, Vigo said in the past decade the fashion industry has pivoted more towards video and multimedia content. “The fashion industry is all about TikTok, Instagram, right? ChatGPT and other AI models are trained on text.”
He added that brands that continued publishing text-heavy content fared better. “Companies that kept paying attention to their blogs, that were frequently featured on text media, those were the ones that managed to sustain their visibility,” he said.
Final gatekeeper
While many chatbots can pull in real-time search results, Vigo said that does not fully solve the problem. “The final gatekeeper is what the model believes,” he said. “If the model hasn’t seen a brand, it doesn’t matter that the brand gets surfaced through the SEO pipeline.”
He likened the issue to advertising against entrenched brands. “If I want to compete with Coca-Cola with a brand called Francisco Cola, and I pay Google millions of dollars to rank number one (in search) when someone enters the keyword ‘Coca-Cola’,” he said, “the model will never mention Francisco as a leading Coca-Cola brand, because it hasn’t seen my brand in the training dataset.”
Vigo said the same dynamics apply to Google’s AI Overviews. “Both Gemini and ChatGPT use RAG pipelines to enrich their solutions,” he said, referring to Retrieval-Augmented Generation. Like ChatGPT, Gemini will “ignore” a brand if it has not seen it in its training data. “We see this interaction between the model memory and the SEO pipeline.”
According to Vigo, geoSurge addresses the problem through what he calls corpus engineering – testing how text data influences model behavior. “We can create different alterations of the corpus, retrain open-source models and see – when we ask the same question to a model that has been exposed to our corpus – if the model responds differently,” Vigo said.
He demonstrated how small prompt changes can dramatically alter responses. Asking an AI model to list the top search engines in the world, Google tops the list, followed by Bing. But tweak the prompt to search for the top search engines for “privacy nerds,” the results are dramatically different. DuckDuckGo topped the list, followed by Startpage, Vigo said.
Those insights, Vigo said, let marketers identify what he calls “perplexity gaps” – areas where models are uncertain and where brands may be able to shape perception. “This is an incredible insight for marketers that want to understand what is their perception from the AI model,” he said.
Go back to text
Vigo also cautioned brands against relying on generative AI to mass-produce content. “One of the biggest mistakes that we see is people trying to use gen AI to generate blog posts for SEO,” he said, citing research that suggests training on synthetic data can degrade model performance.
Despite widespread speculation, Vigo said SEO is not dead. “We see that around 80% of GPT queries are still using search,” he said. However, “SEO is changing and adapting.”
He also warned that many companies underestimate their AI exposure by relying on traditional analytics. “If your brand has been mentioned organically on the text without a link, what marketers are actually doing is they are underestimating their AI traffic and overestimating their organic.”
Asked for one practical takeaway, Vigo was candid. “The piece of advice that I would advise is to not use gen AI to generate content,” he said. “For brands that have been very heavy towards (multi)media, it is now time to start paying attention back to text.”
On image generation, however, Vigo said brands face fewer risks. “I don’t see a very direct relationship between being visible and ChatGPT on the text and image generation. Image generation is a capability of ChatGPT,” he said. “Generating images enables you to experiment at scale and then A-B test to see which has the highest click-through rate.”
As AI-driven discovery accelerates, Vigo said brands must rethink how they measure visibility in a world where models, not links, increasingly decide what users see. “It’s time to go back to language,” he said.
Be First to Comment