A lot of AI search advice circulating right now sounds confident and tests out flimsy. A recent Ahrefs research effort spanning more than 50 studies and roughly 15 million data points set out to check the claims the industry keeps repeating — about schema markup, llms.txt files, backlinks, “best of” listicles, and whether AI search has actually replaced Google. Several widely repeated assumptions did not survive contact with the data.
Here are nine of the findings worth knowing, and what they change about how a site should actually prioritize its AI search work.
Publishing your own “best of” list doesn’t guarantee you’ll be the one recommended
The assumption: publish a “best X” list with your brand at the top, and AI models will recommend you. ChatGPT does lean heavily on this content type — one analysis of 750 ChatGPT prompts found “best X” articles made up 43.8% of all cited page types, which is why so many B2B companies now publish self-ranked lists.
But a controlled Ahrefs experiment publishing 34 self-promotional lists across five domains, then tracking 9,886 answers across ChatGPT, Gemini, Perplexity and Copilot, found that citation and recommendation are two different things. Models frequently used the self-published list as a source without naming the brand that wrote it — and in one case, a list promoting an Ahrefs-run conference resulted in a competing event being recommended in 43% of the AI answers generated from it.
What tends to work better is earning mentions across independent sources — reviews, outreach, word-of-mouth, industry coverage — since that’s the kind of corroboration AI models treat as credible over the long run. A single self-authored list rarely carries the same weight.
Almost nothing reads your llms.txt file
The llms.txt proposal — a plain-text file meant to guide AI crawlers to a site’s key content — picked up a following despite Google publicly stating it isn’t required for AI visibility, while separately suggesting sites audit theirs anyway.
Server-log analysis across 137,000 sites using Ahrefs Web Analytics and Bot Analytics found that 28% had published an llms.txt file — and 97% of those files were never fetched by anything, human or bot. Of the fraction that were fetched, 77% of the requests came from SEO audit tools and GEO platforms studying llms.txt adoption, not from AI crawlers actually using it. In effect, the file mostly gets read by other tools built to measure whether anyone reads it.
The better use of that time is making existing content easier to crawl, parse and extract in the first place, rather than maintaining a file with no measurable audience.
Schema markup isn’t a shortcut to AI citation
Schema optimization gets pitched as a quick win for AI visibility. Testing it directly — tracking 1,885 pages that added JSON-LD schema against 4,000 control pages, comparing citation rates before and after — turned up no meaningful uplift in Google AI Mode or ChatGPT after 30 days. AI Overviews citations actually declined by 4.6% over the same period, though both the treated and control groups were already trending down beforehand, so that drop can’t be cleanly attributed to schema either way.
Schema may still matter indirectly: Google’s Knowledge Graph and databases like Wikidata have relied on structured data for years, and a well-defined entity there could plausibly shape how AI models describe a brand elsewhere. But that’s a slow, indirect path — not the fast citation lever it’s often sold as. Marking up Organization and Person data with sameAs links to Wikipedia, Wikidata and Crunchbase is worth doing for that long-term entity-building reason, just not as a shortcut.
Ranking in classic search still drives most ChatGPT citations
An analysis of 1.4 million ChatGPT prompts, categorizing every retrieved URL by source type, found that 88.46% of citations traced back to the general search index — ordinary organic rankings — rather than specialized channels. Reddit and YouTube get pulled into ChatGPT’s retrieval at scale, but they’re cited far less often relative to how frequently they’re retrieved.
The practical takeaway is unglamorous: classic SEO fundamentals — targeting the right keywords, matching intent, fixing technical issues that suppress rankings — are still doing most of the work behind ChatGPT citations, whatever the retrieval-augmented-generation narrative implies.
Page-one rankings don’t guarantee AI visibility
The flip side is that ranking well doesn’t guarantee showing up in an AI Overview. Ahrefs’ Brand Radar data across 863,000 SERPs and 4 million citations found that only about 38% of URLs cited in AI Overviews also ranked in the top 10 for that exact query — down sharply from roughly 76% a year earlier. The rest ranked lower, or weren’t in the top 100 at all for that specific term. That’s consistent with Google’s own emphasis on query fan-out, where citations increasingly draw from related searches rather than the literal query typed in — a dynamic we cover in more depth in how search users have moved past keywords faster than most content strategies have.
The fix is building topical coverage across a cluster of related queries, not just the one head term a page is nominally targeting.
AI answers reshuffle their sources constantly — but not their underlying opinion
Tracking 43,000 keywords over a month, checking each AI Overview more than 16 times, found the wording changed 70% of the time, the brands mentioned shifted 46% of the time, and the cited sources swapped 45.5% of the time. That volatility makes AI tracking look unreliable at first glance.
But when the comparisons measured meaning rather than exact wording, AI Overviews scored 0.95 out of 1 for consistency. Google has effectively settled on what the answer is; it just keeps re-sourcing and rewording the delivery. Brands that win the underlying topic — rather than chasing exact phrasing — show up more consistently than the surface-level churn suggests. That argues for tracking AI visibility by topic and sentiment over time rather than reacting to day-to-day wording changes, a point that also comes up in why Google’s own AI search reporting can mislead marketers who read it too literally.
Google still sends roughly 190 times more traffic than ChatGPT
Search’s obituary gets written every year, and AI search has given that habit new fuel. The data doesn’t support it yet. Once non-search usage like coding help and translation is stripped out, ChatGPT’s search volume for queries people would traditionally type into Google comes to only around 12% of Google’s. Across 76,000 sites connected to Ahrefs Web Analytics, Google accounted for nearly 40% of site traffic versus roughly 0.21% from ChatGPT — a gap of about 190x. Similar findings elsewhere have shown AI search layering on top of Google’s traffic rather than replacing it.
Google remains the largest channel by a wide margin. The right move is building AI visibility alongside that channel, not instead of it.
An AI Overview appearing above your listing costs more than half your clicks
Google’s own 2024 messaging claimed that links included in AI Overviews get more clicks than the same page would as a standard listing. Independent testing found the opposite. An early-2025 study found AI Overviews reduced clicks to top-ranking content by 34.5%; a repeat of that study in 2026 found the impact had worsened to a 58% cut in position-one click-through rate. Separate analyses from other researchers put the decline anywhere from roughly 47% to over 65%, depending on methodology — directionally the same story despite different numbers. For a closer look at how Google frames these numbers, see Google’s own AI search click figures, and what’s missing from them.
Diversifying traffic sources and tracking actual click loss directly — rather than taking published claims at face value — is the more reliable approach.
Brand mentions matter more than backlinks for AI visibility
Domain Rating and backlink counts are strong predictors of classic search rankings, so it’s a reasonable assumption they’d matter similarly for AI visibility. A study of 75,000 brands, correlating a range of metrics against how often each was mentioned in ChatGPT, AI Mode and AI Overviews, found Domain Rating correlated weakly with AI mentions (0.266–0.326) and backlink counts barely registered at all (0.191–0.244). YouTube mentions correlated far more strongly (around 0.737), with branded web mentions close behind (0.656–0.709).
Links build authority that classic search algorithms reward. For AI visibility, what appears to matter more is whether people are actually talking about a brand — on YouTube, in press coverage, across forums — regardless of how many sites link to it.
Fundamentals beat shortcuts
Across roughly 15 million data points, the pattern that keeps showing up is that the tactics marketed as shortcuts — a quick schema tag, a self-published “best of” list, an llms.txt file — mostly don’t move the needle. What does work is slower and more familiar: ranking well, earning genuine third-party mentions, covering a topic thoroughly rather than narrowly, and building a brand people actually discuss outside a company’s own content. AI search rewards the same fundamentals search always has; it’s just far less forgiving of the tactics built to skip them.
Frequently asked questions
Does schema markup improve AI search citations?
Not directly, based on controlled testing. Adding JSON-LD schema to pages showed no meaningful citation uplift in Google AI Mode or ChatGPT after 30 days. Schema may still contribute indirectly by strengthening entity data in the Knowledge Graph over time, but it isn’t a fast citation shortcut.
Is an llms.txt file worth publishing?
Current data suggests no. Across 137,000 sites studied, 97% of published llms.txt files were never fetched by anything, and most of the fetches that did occur came from audit tools rather than AI crawlers actually using the file.
Do backlinks still matter for AI search visibility?
They matter far less than they do for classic rankings. Domain Rating and backlink volume showed weak correlation with AI brand mentions, while YouTube mentions and general web mentions correlated much more strongly.
Has AI search replaced Google as the main traffic source?
No. Google still sends roughly 190 times more traffic than ChatGPT to the sites studied, even as AI Overviews cut into the click-through rate Google itself sends.