Small publishers lost 60% of search referrals. Google beat a publishers antitrust suit. For brands, the game is becoming a source inside the answer layer.
Google confirmed to The Verge this month that it is running an experiment in Search that uses AI to rewrite article headlines and website titles in results. The company framed it as small and exploratory. The implications are not.
If Google can replace the headline a publisher wrote with one its AI generated, the title tag — the single most visible piece of packaging between your content and the searcher — is no longer yours to control. For anyone building content strategy around AI search, that changes the math on where meaning needs to live inside a page.
What Google Is Actually Testing
The Verge's report documented specific examples of AI-generated headlines distorting the meaning of the original. One article titled "I used the 'cheat on everything' AI tool and it didn't help me cheat on anything" — a skeptical review — was rewritten to simply "'Cheat on everything' AI tool," stripping the skepticism and reading like an endorsement. Another headline about Microsoft "rebranding Copilot in the most Microsoft way possible" became "Copilot Changes: Marketing Teams at it Again."
Google told The Verge: "If we were to actually launch something based on this experiment, it would not be using a generative model and we would not be creating headlines with gen AI." That statement is worth parsing carefully. It does not say Google will not launch AI-altered headlines. It says the launch version would use a different method.
9to5Google independently confirmed additional examples of the test, noting that the AI-generated titles appeared without clear labeling that would distinguish them from the publisher's original.
The Precedent Is Already Set
This is not Google's first attempt at rewriting what publishers put in front of searchers. It is the third escalation in a clear pattern.
2021: Title tag rewriting. In August 2021, Google overhauled its title generation system, pulling title text from multiple on-page signals rather than always using the publisher's <title> element. Early data showed a 77% increase in Google not using the publisher's title tag on the SERP. But the key constraint was that Google drew from existing text on the page — it assembled, it did not invent.
2025: AI headlines in Discover. Google tested AI-generated headlines in its Discover feed in mid-2025. Despite documented accuracy problems — including a headline falsely claiming a Steam Machine price had been revealed — Google made the feature permanent in January 2026, stating it "performs well for user satisfaction." The lifecycle from experiment to permanent was roughly four weeks.
2026: AI headlines in Search. The current test extends the same approach to the core search results page. The direction of travel is obvious.
Why This Matters for AI Search Strategy
Most content strategy still treats the title tag as the primary packaging layer. It is what shows up in search results, browser tabs, social shares, and link previews. Publishers craft it carefully because it is the first — and often only — piece of content a potential reader sees before deciding to click.
If Google's AI can replace it with whatever its model generates, that packaging layer becomes unreliable. The meaning of your page must not depend on the headline to be understood correctly.
This is not hypothetical. Google's own documentation on title links lists seven signals it already draws from when generating titles: the <title> element, the main visual heading, <h1> tags, og:title, prominent styled text, anchor text, and WebSite structured data. The AI experiment goes further by generating entirely new text.
Build Meaning Into the Body, Not Just the Headline
The practical response is to make your content self-interpreting at the chunk level. A few specific shifts:
Lead with definitional framing. The first paragraph of every page should state what the content is about in plain, extractable terms. Do not rely on the headline to provide context that the body does not repeat.
Make H2s descriptive, not clever. Subheadings that describe the content of their section ("How AI Overviews Select Sources" vs. "The Big Picture") serve as fallback framing if the title gets rewritten or stripped.
Front-load key claims. Research on how language models process long contexts — particularly the "Lost in the Middle" paper published in the Transactions of the ACL — shows that LLMs perform best when relevant information appears at the beginning or end of input. If an AI system processes your page in chunks, the opening paragraph and first heading carry disproportionate weight.
Use structured data as a semantic anchor. Your Article schemaheadline and description properties give AI systems a machine-readable statement of what the page is. That structured signal may prove more durable than the visual title tag as Google continues to experiment with presentation-layer overrides.
The Bigger Pattern: Platform Control Over Presentation
This headline experiment does not exist in isolation. It is part of a broader shift where platforms exercise more control over how publisher content is presented to users.
AI Overviews already reframe and synthesize page content without preserving the publisher's original structure or emphasis. Digital Content Next's survey of 19 major publisher companies found median year-over-year referral traffic from Google down 10%, with some news brands seeing weekly declines of 16%. The European Commission opened a formal antitrust investigation in December 2025, examining whether Google imposed unfair terms on publishers while using their content for AI purposes.
The headline rewriting experiment is the logical next step. First the platform summarizes your content. Then it repackages the snippet. Then it replaces the headline. Each step moves the publisher further from the presentation layer.
The Takeaway
If you are building content for AI search visibility, the title tag was already one of many signals rather than the signal. This experiment makes that even more explicit. The headline is a suggestion, not a guarantee. The page body, opening definitions, and chunk-level framing are where retrievable meaning must live.
Build every page as if the headline might not survive contact with the platform. Because increasingly, it will not.
James Calder is the editor of The Search Signal, covering AI-powered search, generative engine optimization, and the future of brand discovery.
Small publishers lost 60% of search referrals. Google beat a publishers antitrust suit. For brands, the game is becoming a source inside the answer layer.
Perplexity's Personal Computer turns a Mac mini into a persistent, always-on AI agent that works across your local files and apps around the clock — a new category of ambient computing.
Brave Search now offers two AI-native API endpoints that deliver pre-extracted, relevance-scored web content directly to LLMs — signaling a fundamental shift in how search APIs serve machine consumers.
