Google swapped the model behind AI Mode. The links got quieter.

Google put a new model into AI Mode on September 2, 2026. By the next morning, experienced SEOs were posting AI Mode answers that carried no links and no citations at all.

The easy read is "Google is killing links." The more useful read is sitting in Google's own documentation, and it was there before this week: which model is answering changes which links you see — and Google describes web links, in part, as what you get when the model is not confident.

Sources: Introducing Gemini 3.8 Flash and 3.8 Flash Cyber, Google, September 2, 2026; Google Search Using Gemini 3.8 Flash In AI Mode, Search Engine Roundtable, September 2, 2026; Google AI Mode Gemini 3.8 Flash Not Link / Citation Friendly?, Barry Schwartz, Search Engine Roundtable, September 3, 2026; AI features and your website, Google Search Central (last updated 2025-12-10 UTC); Get AI-powered responses with AI Mode in Google Search, Google Search Help.

The short version

• Google's post, September 2: 3.8 Flash "is available to Google AI Pro and Ultra subscribers across the Gemini app, AI Mode in Google Search and Gemini in Google Sheets."
• That post says nothing about links, citations, or supporting sources. Not one line.
• September 3, Search Engine Roundtable: early feedback is that AI Mode on the new model "is less likely to cite sources and link to sources than previous models." Named practitioners, screenshots, top-of-funnel queries.
• That is observation. Not a measurement, not a rate, and not a Google statement.
• Google's AI features guide already said the variance out loud: "AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary."
• Google's AI Mode help page describes links partly as a confidence fallback: "In some cases, AI Mode will provide a set of web links if there's not high enough confidence in the quality or helpfulness of an AI response."
• Eligibility did not change: "a page must be indexed and eligible to be shown in Google Search with a snippet." No new file, no special markup.
• If a model release can take your citation away in a day, the citation was never the asset. The conditions under it are.

What Google actually shipped

Google's September 2 post introduces Gemini 3.8 Flash and a cybersecurity variant. The line that matters to a site owner is in the get-started list:

Consumers: 3.8 Flash is available to Google AI Pro and Ultra subscribers across the Gemini app, AI Mode in Google Search and Gemini in Google Sheets.

Two facts, and only two. The model is in AI Mode. It is there for paid subscriber tiers, not for everyone landing on Search.

Google also describes the behaviour change in general terms:

These performance gains stem from a core design choice: 3.8 Flash works harder. On complex tasks, it exhibits greater diligence — executing extra reasoning steps, and calling tools iteratively.

That is a reasoning claim. The post does not mention links, citations, or supporting sources anywhere.

Search Engine Roundtable covered the rollout the same day, quoting Google Search VP of Product Robby Stein on X — "Gemini 3.8 Flash has landed in Search" — and, in a second quote, "Available today in AI Mode for Google AI Pro & Ultra subs around the world." Search Engine Land ran the same rollout on September 2. Neither Google statement says anything about how many sources an answer carries.

For the record: there was no Google Search Central blog post on September 1, 2 or 3. The newest post there is the August 28 update to the site reputation policy.

What operators actually saw

On September 3, Search Engine Roundtable ran the observation:

As you know, Google rolled out Gemini 3.8 Flash for subscribers to AI Mode and some have started to play with it. And the early feedback is that AI Mode running Gemini 3.8 Flash is not link or citation friendly.

The specifics from that post: Gagan Ghotra posted a run of top-of-funnel AI Mode queries where none of the answers showed links or citations. Glenn Gabe posted a side-by-side of the same query on 3.8 Flash and on the default model, tagging Google product leads. Barry Schwartz is candid about the state of it — he is not sure why it would be, and hopes Google can fix it soon.

Take that for what it is worth, which is real but small. A handful of experienced people, running queries on a subscriber tier, on day one of a rollout, posting screenshots. Worth annotating in your log. Not a study, not a rate, and nobody has published a defensible before-and-after on a fixed prompt set.

Anyone handing you a precise "citations dropped X%" figure this week is selling you a chart, not a finding.

Google already documented the mechanism — twice

This is the part the panic threads skipped.

One: the link set moves with the model. From Google's AI features guide:

AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary.

That sentence has been in the documentation since last year. A model swap changing the link set is not a broken promise. It is the documented behaviour of the surface.

The same guide also carries the sentence this week's screenshots appear to contradict, which is worth quoting exactly rather than pretending it does not exist:

While responses are being generated, our advanced models identify more supporting web pages, allowing us to display a wider and more diverse set of helpful links associated with the response than with a classic web search, enabling new opportunities for exploration.

Descriptive, not a commitment. But if the new model routinely shows no links at all on top-of-funnel queries, that description and the live product have drifted apart, and Google is the only one who can close the gap.

Two: links are partly a low-confidence behaviour. From Google's AI Mode help page:

In some cases, AI Mode will provide a set of web links if there's not high enough confidence in the quality or helpfulness of an AI response.

Put those next to a launch post that says the new model "works harder" and "exhibits greater diligence," and there is an unglamorous hypothesis that fits everything on the record: a model that is more confident in its own answer surfaces fewer fallback links.

That is our read, not Google's statement. Google has not said it. But it is a better working theory than "Google decided to punish publishers on a Wednesday," and it points at a different fix. If the link slot narrows to the queries where the model doubts itself, then the pages that keep getting pulled in are the ones carrying specifics a model cannot confidently generate from memory: your prices in text, your service areas, your hours, your named people, your dated proof.

Worth noting on the same page: Google's AI Mode help page still describes the model menu as "Fast" and "Pro," and names Gemini 3 Pro. It has not been updated to mention 3.8 Flash. The documentation is behind the rollout, which is its own reason not to treat day-one behaviour as settled policy.

What did not change

The eligibility floor:

To be eligible to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements. There are no additional technical requirements.

The absence of a guarantee:

Just because a page meets all requirements, best practices, and complies with the policies, doesn't mean that Google will crawl, index, or serve its content. Indexing and serving isn't guaranteed.

And the answer to everyone still selling a file as a citation strategy:

You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add.

So: the bar to be eligible did not move this week. The answer style on one tier of one surface may have. Those are different problems and they get different responses.

Why this is the whole argument

Our position has not changed, and this week is why we hold it. You cannot buy, earn, or optimise your way to a guaranteed citation slot, because the slot is not yours. It belongs to a model that can be replaced on a Tuesday.

What you can own is the set of conditions underneath it. Indexed. Crawlable. Snippet-eligible. First-party answers actually worth quoting. Named authors. Prices, service areas and offers stated in text a machine does not have to guess at. Evidence a buyer can inspect.

Those survive a model swap. A citation does not.

It is also the honest limit of every AI-visibility dashboard on the market, including the good ones. Monitoring tells you the answer changed. It cannot tell you the model changed underneath you, and it cannot ship the page that earns you back in. Monitoring is not execution.

Same week, same lesson

Two other items landed inside the same 24 hours, and they rhyme.

Publishers told Search Engine Roundtable on September 3 that content being crawled, indexed and served in Google Search results was going missing from the News tab, sometimes reappearing much later, with no pattern anyone could find. Separately, Google's Merchant Center help page now reads: "AI performance insights are currently available for English-language queries for Merchant Center accounts in Australia, Canada, India, New Zealand, and the United States" — a measurement surface that started with the United States and is widening.

Different surfaces, one lesson: presence on one Google surface is not presence on all of them, and each surface gets its own reporting on its own schedule. Track them separately or you will read a surface-level change as a site-level penalty.

The 7-day operator plan

1. Freeze the diagnosis. Do not tear up a content plan on day one of a model rollout. Nothing in Google's eligibility documentation changed this week.
2. Baseline a fixed prompt set today. Twenty to forty buyer questions you actually care about, same wording every time, dated screenshots. Record whether links appeared at all and which domains got them. Without a baseline you cannot claim you lost anything.
3. Write down the tier and the account state. The new model went to Google AI Pro and Ultra subscribers. A signed-out spot check is not the same surface your prospect is on. Note which state you tested from, every time.
4. Separate the three failure modes. Not indexed. Indexed but not snippet-eligible. Eligible but not selected. Only the third is a model-behaviour story. The first two are your own site and are fixable this week.
5. Check the floor, not the vibes. Confirm the pages you want cited are indexed, crawlable in robots.txt and at the CDN or WAF, snippet-eligible with no stray nosnippet or over-tight max-snippet, with the important content in text.
6. Write the specifics a model cannot invent. If links are surfacing where the model is less confident, the pages that keep earning them are the ones carrying facts only you hold: current pricing in text, service-area lists, staff names and credentials, dated results. Generic explainer prose is exactly what a confident model replaces.
7. Keep Search Console on its own sheet. Google's guide says sites appearing in AI features are included in overall search traffic in Search Console, reported in the Performance report within the "Web" search type. Do not invent an "AI traffic" number the reports do not give you.

Honesty block — what nobody confirmed

• Google did not announce any change to AI Mode links or citations. The September 2 post is a model launch. Links, citations and supporting sources are not mentioned in it.
• We found no Google response to the missing-citations reports as of this writing on September 3, 2026 — not from Search Liaison, not from the product leads tagged in those threads. Google's Search Status Dashboard showed crawling, indexing, ranking and serving all clear through September 3, with no AI Mode incident logged. Worth noting that when SEOs flagged an AI Mode citation bug in April 2026, Search Engine Roundtable reported a Google product lead replying within a day that it was a bug. The silence this time is part of the story, not evidence of anything.
• No rate, no percentage. We are not publishing a citation-rate figure. The public evidence is day-one screenshots. Vendor posts circulating with precise citation-decline percentages are not sourced to anything we could open and verify, so they are not cited here.
• The confidence-fallback reading is ours, not Google's. Google's help page says links appear "in some cases" when confidence is not high enough. Google has not connected that to this week's behaviour. Treat it as a working hypothesis that fits the record, and drop it the moment Google says otherwise.
• Not universal, and possibly not permanent. This is reported on paid subscriber tiers, on top-of-funnel queries, in the first day. Rollouts get tuned, and Google's own AI Mode help page has not even been updated to name the new model yet.
• We did not test this at scale ourselves. This post reports what Google published and what named operators posted publicly. When we have a dated before-and-after on our own fixed prompt set, that will be its own post, with the method attached.
• No promise in either direction. We cannot promise AI will mention you, and we are not going to promise a model swap will hand your links back.

Related reading

• ChatGPT can now use tools on your page. Getting cited did not buy you that.
• Google AI Mode Can Call Your Business: What to Fix Now
• A 1,120-Query Study Found a 28.5% Local Pack-to-AI Mode Visibility Gap
• GEO measurement is noisy, so build a proof loop
• The GPT-5.6 Rumors Are the Wrong Question
• Why Good Brands Disappear From AI Search

In-body CTA

I am not going to pretend a scan is a strategy. It is a free check — email only, no card — that tells you whether your pages clear the floor every one of these surfaces starts from, before you spend a week arguing about a model you do not control.

Get your AI-visibility score