September’s State of Search & AI covers why AI recommends brands it already knows from training, what ChatGPT’s index actually stores from your pages, and more.
September’s State of Search & AI covers why AI recommends brands it already knows from training, what ChatGPT’s index actually stores from your pages, and more.
Most AI search optimization work focuses on what happens after a query is submitted. Can the model find your pages, parse them, and decide to cite them?
September research suggests that, for recommendation queries specifically, much of the decision has already been made. Models carry associations between brands and categories from training — and those associations are shaped more by sustained web presence than by technical fixes. That finding sits alongside a specific and actionable discovery about how much AI actually indexes from a given page, and some encouraging data on what AI-referred traffic does when it arrives.
August drew a sharp line between citation and recommendation. Brands cited as a source inside an AI Overview were excluded from the model’s actual recommendation 69% of the time. A large-scale keyword study added nuance to the zero-click narrative, showing informational categories are taking the heaviest losses. And independent testing found AI tools were regularly surfacing wrong addresses and postcodes for multi-location retailers.
September gets more specific about why AI recommends some brands and not others — and what’s within a marketer’s control.
This month covers model updates, a legal development with broader data implications, a technical finding about how much AI indexes per page, and traffic data that reframes the AI-versus-organic question.
Google rolled Gemini 3.7 Flash into AI Mode one day after its public release, available to paid subscribers via the model picker. Google describes it as “better at following instructions and understanding intent,” which typically signals changes in how the model handles ambiguous or layered queries, and by extension, which pages it pulls from. The cadence is worth noting: Gemini 3.5 Flash in July, 3.7 Flash in September. Model updates are becoming a regular variable in AI Mode performance.
What this means for marketers: Monitor AI Mode citation and impression patterns across model updates the way you’d monitor organic performance across algorithm updates. Establishing baselines now makes meaningful shifts detectable when they happen.
Testing of ChatGPT’s recommendation behavior found that for category-level queries (ex: what’s the best brand for X), the model leans heavily on associations formed during training rather than live web retrieval. Brands not already connected to their category in training data had roughly a 2% chance of being recommended. Recommendation equity grows over time as a brand is widely reviewed, referenced, and discussed across the web, in content the model was trained on.
What this means for marketers: Technical SEO and page optimization matter for factual and citation queries, but category-level AI recommendations are driven by a different signal. The brands that earn those recommendations have a deeper footprint across earned media, PR, and third-party coverage.
Google refiled its DMCA lawsuit against SerpApi after an initial dismissal, narrowing the claim to licensed content from partners like Reddit who required Google to block unauthorized scraping. SerpApi’s data feeds power a significant share of the rank-tracking and SERP feature reporting tools the SEO industry runs on. Even a narrow Google win could make that data harder or more expensive to obtain across the board.
What this means for marketers: Nothing actionable yet, but worth knowing which reporting tools in your stack rely on scraped SERP data. First-party data from Search Console and GA4 becomes more valuable if third-party access is limited.
OpenAI clarified that its ChatGPT-User bot, which fetches a page live when a user asks about a specific URL, may not honor robots.txt disallow rules, categorizing those fetches as user-initiated rather than automated crawling. TollBit’s H1 2026 data found that roughly 15% of AI page fetches on tracked European sites reached disallowed pages. Anthropic states all three of its bots honor robots.txt consistently.
What this means for marketers: For content that truly needs to stay out of AI responses (internal tools, staging environments, draft pages), robots.txt isn’t a reliable barrier against ChatGPT’s fetch bot. That content needs to sit behind authentication.
Shopify’s Q2 data found AI-referred storefront sessions grew 197% year-over-year, converting at roughly twice the rate of organic in research-heavy categories. Organic search still drove more total traffic than all AI platforms combined. The data also reinforced that structured product data improves performance across both channels.
What this means for marketers: The AI-versus-organic framing is the wrong frame. Both channels are growing, and the investments that benefit one largely benefit the other.
Analysis of ChatGPT’s free-tier index found it stores only a page’s title and approximately 200 characters of body content, which is about two sentences. Any boilerplate, generic introductions, or navigation text that appears before the first substantive paragraph eats into that space.
What this means for marketers: On pages where AI citation matters — location pages, product pages, key service pages — the first sentence of real substance should appear as early in the body as possible. Page templates that open with a brand statement or disclaimer before the actual answer are using the most valuable index space on content that won’t get cited.
Citation and recommendation continue to be treated as separate problems with distinct solutions. Citation is an optimization problem targeted with crawlability, content structure, and placing important copy as early on the page as possible. Recommendation is a brand-building problem focused on the depth of a brand’s presence across the web content that models trained on, built over time through earned media, PR, and third-party coverage. Most AI search strategies are narrowed in on the first and underinvesting in the second.
For marketers, that means:
AI recommends who it already knows. Building that familiarity across the web and over time is the work that makes everything else worth doing.
September’s State of Search & AI covers why AI recommends brands it already knows from training, what ChatGPT’s index actually stores from your pages, and more.
Stop following the conversation — lead it. This guide breaks down how to build a thought leadership content strategy that earns citations and real authority.
Got gaps in your content funnel? See how to connect awareness, consideration, and decision content so prospects can follow the crumbs straight back to you.
