At the start of the year, we predicted that 2026 would be the year AI moved from a supplementary research tool to the foundation of how buyers form shortlists. Six months in, the prediction looks conservative. What we did not fully anticipate was how much the AI systems themselves would change underneath that shift, and how differently each of them would come to see the same market.
This is a mid-year review in three parts: what changed in the models, what changed in AI search, and what our own data across more than a hundred categories says about how those changes are playing out in the recommendations buyers actually receive.
Part One: The Models Turned Over Again
Every major lab replaced or substantially updated its flagship in the first half of 2026. OpenAI's GPT-5 line moved through successive point releases. Anthropic shipped the Claude 4.6 generation early in the year, followed by further Opus updates. Google's Gemini 3 family entered preview while the 2.5 models remained the production workhorses. xAI released Grok 4.20 in March. DeepSeek previewed its V4 architecture in April. Perplexity's Sonar models continued to iterate on top of the retrieval layer that makes them different from everything else on the list.
For a brand, the significance is not any single release. It is the cadence. The systems that recommend vendors to buyers are now replaced or retrained every few months, and each replacement carries a different training cutoff, a different weighting of sources, and a different picture of your market. A position earned with one model version is not guaranteed with the next.
QuadrantX updated its own production recipe at the end of April to Claude Sonnet 4.6, GPT-5, Gemini 2.5 Flash, DeepSeek, and Sonar Pro, precisely so that our scores reflect the systems buyers are using now rather than the ones they used last year. Our historical data lets us see what such a switch does: in several categories, a new model generation reordered the shortlist without any change in the underlying market.
Model Turnover Risk — The exposure a brand has to changes in its AI visibility that are caused by AI providers replacing or retraining their models, rather than by anything the brand or its competitors did. It is measurable only with a historical, multi-model record, and it is one of the reasons single-point-in-time AI visibility checks mislead.
Part Two: AI Search Stopped Agreeing With Itself
The second half of 2025 and the first half of 2026 produced a body of measurement about how AI answers choose their sources, and its central finding is that the old relationship between search ranking and AI citation has broken down.
- Ahrefs found that the share of Google AI Overview citations drawn from top-ten organic results fell from 76% in July 2025 to 38% by March 2026, as Google's query fan-out pulled citations from far deeper in the index.
- The same firm found that Google's own AI Mode and AI Overviews cite the same URL for a given query only 13% of the time. Two AI surfaces from one company, answering one question, rarely agree on the evidence.
- Semrush's three-month study of 230,000 prompts found the most-cited domains across ChatGPT, AI Mode, and Perplexity shifted sharply month to month, with Reddit's share of ChatGPT citations collapsing from around 60% to around 10% within a fortnight in September 2025.
- Cloudflare's measurement of crawler behaviour showed AI companies fetching thousands of pages for every visitor they refer, which prompted publishers to start blocking AI crawlers and, in turn, began to change what the models can see.
The practical consequence is that "rank well and the AI will cite you" is no longer a reliable strategy. Each AI surface has its own source diet, and that diet changes. The teams doing well in our data are the ones that stopped optimising for one engine and started measuring across several.
One more development deserves mention because it answers the question every marketing leader asks. Similarweb's clickstream analysis, published at the end of June, found that when AI assistants recommended a financial brand, users were measurably more likely to visit that brand's site directly in the following week, with lifts of 7% to 14% for the brands studied, and a corresponding drop in branded search. AI recommendations are not a vanity metric. They move traffic, and they move it around the search engine.
The AI search market of mid-2026 is not one answer engine. It is a dozen, each with its own sources, its own biases, and its own idea of who leads your category.
Part Three: What Five Models Say About One Market
QuadrantX asks five production AI models the same category questions, on a rolling schedule, across more than a hundred categories. That gives us something the industry studies above cannot: a direct, repeated comparison of how different models recommend vendors when asked the same thing. Three findings define the mid-year picture.
Agreement is the exception
All five models named the same market leader in 38% of categories. At least four agreed in 62%. In 15% of categories no majority existed at all. Below the leader, a typical category produced about ten distinct vendors across the five models' top-five lists, twice what perfect agreement would produce, and 41% of all vendors mentioned were mentioned by one model only. The market a buyer sees depends on which assistant they open.
Each model has a personality
The disagreement is not random. GPT-5 names the most vendors per category, around 27 against 20 to 22 for the others, and produces the widest long tail. Perplexity's Sonar Pro, the only model in our set that searches the live web before answering, diverges from the group's leader in about a third of categories, more than any other, and its divergent picks are consistently the most recent and most specialised. Claude Sonnet 4.6 diverges least. Gemini 2.5 Flash and Sonar Pro score vendors several points higher on Sentiment than DeepSeek does. These tendencies are stable across categories and across editions, which means they are properties of the models, not noise.
Consensus, once formed, holds
In roughly half of our categories the leader has held its position in more than nine editions out of ten this year. Salesforce in CRM, AWS in cloud infrastructure, McKinsey in management consulting, and State Farm in Texas insurance have not been displaced by any model in any edition. In the other half, the lead moves, sometimes in most consecutive editions. The categories that move are the ones where models disagree, and they are disproportionately the specialist, regional, and newly defined markets where the public narrative is still thin.
For the full analysis, see why consensus scoring produces more reliable market intelligence.
What Is Coming
Four things to expect in the second half of 2026, based on the trajectory of the first.
More turnover. Every lab has signalled further releases. Each one will reshuffle some shortlists. Brands that measure continuously will notice; brands that check annually will attribute the change to the wrong cause.
Retrieval everywhere. The distinction between training-data models and retrieval-based models is blurring as more assistants search the web before answering. In our data, retrieval is where new content shows up first and where specialist vendors break through. As retrieval spreads, the reward for fresh, specific, citable content grows.
A source fight. Publishers blocking crawlers, licensing deals shaping what models can cite, and AI companies negotiating access will change the source diet of every model, unevenly. Some brands will lose visibility for reasons that have nothing to do with their market position. Only a multi-model, historical record will distinguish a crawl-access problem from a narrative problem.
Commercial pressure on answers. Advertising and shopping features inside AI assistants are arriving. The integrity of organic AI recommendations, and the ability to measure them separately from paid placement, will become a live issue for anyone using AI visibility as a business metric. QuadrantX scores are built from organic model responses only, and we intend to keep it that way.
The first half of 2026 established three facts: the models change every few months, the sources they cite are diverging rather than converging, and the models disagree with each other about most markets below the very top. The response to all three is the same. Measure across models, measure over time, and treat the differences between models as the signal rather than the noise.
The Bottom Line
The AI landscape of mid-2026 is larger, faster-moving, and less internally consistent than the one we described in January. That is uncomfortable for anyone hoping for a single number that captures their AI visibility. It is also the reason a multi-model, longitudinal view has become the minimum standard for market intelligence. You can explore how the five models see any category we track on the QuadrantX Explore page, or read our earlier guide to the major AI models for the background.