For a year, the story about AI search has been fragmentation: a new assistant every month; a browser that shops; a spreadsheet that reasons; a phone assistant that finally works. The instinct is to treat each one as a new place you have to go win. This week gave us two developments that argue for the opposite instinct and, unusually, some hard evidence about what winning actually requires.
The consolidation stopped being a theory
Start with a release note that most marketers may have scrolled past. Microsoft's July update to Microsoft 365 Copilot added two frontier models inside Excel: OpenAI's GPT-5.6 and Anthropic's Claude Opus 5. Opus 5 had launched only days earlier - Anthropic's fourth Claude 5 model in under two months, and it had already surfaced inside Perplexity's enterprise product too.
So, in a single week, one model (Claude Opus 5) showed up composing answers inside Microsoft Copilot and inside Perplexity, two products that look like competitors and market themselves as distinct destinations. The logos multiply, the models underneath consolidate onto a handful of frontier systems. This is the same pattern already visible elsewhere: Google's AI Overviews and AI Mode both run on Gemini, and Apple's Siri routes its harder questions to a custom Gemini model too.
For anyone trying to be recommended by AI, the practical consequence is liberating, not alarming. You do not need a separate, panicked strategy for every new box with a chat cursor. "Getting recommended on Copilot" and "getting recommended on Perplexity" are not two problems, because they substantially share the models doing the retrieving and reasoning. Map the surface to the model before you map the budget, and you stop paying three times to solve one problem.
Now we know more about how those models choose
The harder question has always been: once the same few models are deciding, *what* makes them cite you? This week produced the best answer yet. A study presented at SIGIR 2026, "What Gets Cited: Competitive GEO in AI Answer Engines," ran a genuinely rigorous experiment: 252,000 controlled trials across six large language models, feeding each model exactly two candidate sources and measuring which one earned the first citation. Crucially, the researchers anonymised brands and counterbalanced the order of sources, so they could separate *what the content said* from *which name was on it* and *where it sat in the list*.
The headline finding: the two biggest drivers of being cited first are: topical relevance and list position. Not brand familiarity.
A source that more precisely matched the question, and a source that appeared earlier in the retrieved set, won the citation, repeatedly, across models. The study also confirmed a real "position bias": the same source cited less often when it sat second rather than first.
If we read this alongside the July GEO survey (which concluded that widely-cited "GEO visibility gains" only hold when a source is already present in the model's context), a consistent picture emerges. Being cited is a two-gate problem. First you have to be retrieved into the consideration set at all. Then, among the sources that made it, relevance and placement decide who gets named. Brand fame helps far less than the dashboards imply.
Picture a homeowner with a burst pipe asking the AI for "the best emergency plumber near me." Two gates decide whether your plumbing business gets named. Gate one: does the AI even pull your business into the short stack of options it looks at? If your details aren't out there in a form it can find, you're not in the running. It simply never sees you. Gate two: once you're in that stack, does your listing actually say, in plain words, "24-hour emergency plumber in this town, burst pipes, licensed"? The plumber whose page answers the exact question, sitting near the top of the pile, gets recommended, even over the bigger-name shop down the road whose page just says "plumbing services." Being the most famous plumber in town doesn't win it. Being the clearest, most on-point answer in the stack does.
What to actually do
Two moves follow, and they're the two we make every week because the evidence keeps pointing at them.
First, optimise for models, not logos. Before you treat Copilot, Perplexity, and Siri as three campaigns, understand how much retrieval and reasoning they share. The frontier model composing the answer is the thing that decides whether you appear. The surface just decides where it's displayed.
Second, be the most relevant answer, not the most famous one. The SIGIR result is a permission slip for smaller and challenger brands: you do not have to out-spend an incumbent's brand recognition to get cited. You have to match the buyer's actual question more precisely, in language a model can lift directly, and earn your way into the retrieved set. Relevance and placement are things you can build. Brand fame is not something you can buy your way to overnight. And, it turns out, you may not need to.
The surfaces will keep multiplying. The models behind them are consolidating, and we now have peer-reviewed evidence about what they reward. Match the question, earn the placement, and you get cited on every surface those models power.
---
Clarity Digital Advisory helps B2B companies get recommended, not just mentioned, by the AI platforms buyers now use to shortlist vendors.