Why Google AI Overviews Now Send Links to AI Mode, Not Your Website

Something changed inside Google's AI Overviews this week, and it is small enough to miss and big enough to matter. Links in an AI Overview no longer reliably take the reader to a website. They open another question instead. That was spotted on 21 September, written up by Search Engine Roundtable the next day, and it is the clearest signal yet about where Google wants the conversation to end.

What changed in Google AI Overviews this week?

On 21 September, SEO specialist Gagan Ghotra documented anchor links inside AI Overview answers that behave like normal links but open a follow-up question in AI Mode instead of loading an external page. Search Engine Roundtable confirmed and published the screenshots and video on 22 September. The links look ordinary, sit near the bottom of the answer, and function as suggested follow-ups. Clicking one keeps the reader inside Google's conversational surface. Two honest qualifiers: this is an observed interface test, not a feature Google has announced or rolled out widely, and Google has not commented on it. But it removes the last element of a search result that everyone assumed still meant "go to a website". The link shape survives. The destination changed.

Why does this matter more than the usual Google tweaking?

Because the hyperlink was the one part of the results page that was still honest about intent. Every previous change left the mechanics alone and moved the furniture: AI Overviews sat above the blue links, then got taller, then got a "Show more" button. The mechanics held. Now the mechanics are the thing being changed. That follows a pattern that started at the end of August, when Google confirmed that AI Overviews can load fully expanded for some queries, with the "Ask anything" box already open and no click required. An audit of 50 B2B informational and comparison queries on 10 and 11 September found 64 per cent triggered a pre-expanded Overview, and of the 32 that expanded, only 11 cited a source in the visible top fold. Publishers are working from rough estimates of another 15 to 40 per cent reduction in clicks on informational queries. One mercy exists: if you have already started scrolling, Google cancels the expansion so you do not lose your place.

Did ChatGPT really stop citing Reddit?

Yes, and by a startling margin. Promptwatch data reported through Search Engine Land shows Reddit's share of ChatGPT Search citations fell 86.4 per cent in four days, from 3.83 per cent to an average of 0.52 per cent. Otterly's September dataset measures the same thing from a different panel: ChatGPT's Reddit citations are down about 94 per cent against July, and Microsoft Copilot now cites Reddit at zero. Perplexity held up far better, slipping about 11 per cent to 2.04 per cent. YouTube went the other way entirely, with Perplexity citations up 250 per cent, Copilot up 167 per cent and ChatGPT up 13 per cent across September. If any part of your plan leaned on Reddit threads for AI visibility, this is the week that plan broke, at least for ChatGPT. Diversify across YouTube, your own site and review platforms rather than betting on one source.

Does schema markup actually get you cited?

Far less than most agencies imply. A study tracking 1,885 pages that added JSON-LD structured data between August 2025 and March 2026, measured against 4,000 control pages, found no statistically significant citation lift on AI Mode or ChatGPT, and a small, probably unrelated decline on AI Overviews. A separate three-month experiment reached the same conclusion. Read that carefully rather than throwing your schema away. Structured data remains worthwhile technical hygiene, and it helps Google understand your entity. It is an amplifier, not a driver. What actually moved citation rates in the Princeton and Georgia Tech research was inline citations to primary sources, worth roughly a 40 per cent lift, specific statistics at about 37 per cent, and named expert quotations at about 22 per cent. Third-party trust dominates everything: brands with no active review presence were cited in roughly 1 per cent of relevant answers, against more than 75 per cent for brands that actively collect and respond to reviews.

What does Google say you can ignore?

Google has been unusually direct. Its guidance states that its generative AI features run on the same ranking and quality systems as regular search, which it summarises as optimising for the search experience and therefore still SEO. It then names tactics you can skip for its own AI surfaces: chopping content into tiny chunks, building bespoke schema just for AI, and, for Google Search specifically, a dedicated llms.txt file. The nuance worth holding onto is that Google speaks only for Google. Other assistants read the web on their own terms, so a clean technical setup, clear structure and accessible crawlers still matter across the wider set of engines. Scope the claim to the engine before you spend money on it.

What should a UK small business do this week?

• Check your own money queries by hand, logged out. Not two hundred keywords. The ten or fifteen that actually produce enquiries. Note which ones now open a full AI answer and which still return ordinary results. That split is the most useful map you have, and no tool will hand it to you.
• Write for the second question, not just the first. For every commercial page, ask what someone would type into that follow-up box. Price, comparisons, eligibility, timelines, objections. Answer those on the page instead of burying them in a sales narrative.
• Lead every section with the answer. Question-shaped headings, then a direct, self-contained answer in the first two or three sentences. Detail after. That suits a reader skim and a machine extracting a passage.
• Put real numbers and named sources in. Dated statistics and links to primary sources are what the research shows moving citation rates, and they cost nothing but care.
• Fix your review pipeline before your markup. The gap between no review presence and an active one is the single largest lever in the data.
• Stop reporting AI Overviews and AI Mode separately. They are one surface now.

What can you actually measure now?

Since 31 August, Google's generative AI performance reports in Search Console have been available worldwide. They show how often your URLs appeared inside AI Overviews and AI Mode, split by page, country, device and date. The catch matters: impressions only, with no click data and no visibility into the prompts that triggered them. Google merged the two surfaces into one number, so a single generative visibility figure is now all the first-party data you get. Put it next to enquiries rather than next to last month's version of itself, and remember that this surface churns: roughly four in ten AI citations drop within a month, and seven in ten are gone within six. Trend matters. Snapshots do not.

The short version

Google is turning the AI Overview from a destination into a doorway, and this week it quietly removed the last part of the result that still pointed outwards. At the same time, ChatGPT's appetite for Reddit collapsed and fresh research knocked schema markup off its pedestal. The businesses that stay visible will be the ones that answer the follow-up question on their own page, back it with real evidence, and build trust on the platforms AI actually reads. None of that requires new tools. It requires clearer writing and a review pipeline that works.