Google AI Mode Stopped Citing Sources — Fix Coming

Google has confirmed a bug in AI Mode, its conversational search experience, where answers generated by its newest model, Gemini 3.8 Flash, frequently failed to link to or cite the websites they drew from. Google says a fix is coming, though it has not given a public date. For marketers, the practical read is this: if your Google referral traffic softened recently without any change in your rankings, a missing-citations bug is now a plausible explanation rather than a paranoid one.

The issue was reported by Search Engine Land, which noted that Google acknowledged the problem after the behaviour was spotted in the wild.

First, what AI Mode actually is

AI Mode is the tab in Google Search where you ask a question in plain language and get a written answer rather than a page of blue links. Underneath, it runs on a large language model (LLM) — software trained to predict and generate text — which reads across web pages and assembles a response.

The important design detail is that AI Mode is supposed to show its work. Alongside the generated paragraph, it surfaces links to the pages it used. Those links are the entire economic bargain of AI search: publishers and brands let Google read their content, and in exchange they get attribution and, ideally, a click.

Remove the links and the bargain quietly collapses. The answer still gets delivered. The source just stops existing as far as the reader is concerned.

What actually broke

Google, like every major AI company, ships new model versions continuously. Gemini 3.8 Flash is a “Flash” model — the fast, cheaper tier built to serve enormous query volume at low latency. That is exactly the kind of model you would put behind a search product used by hundreds of millions of people.

The reported problem is that this model version was unlikely to attach citations to its answers. Not that it cited the wrong pages — that it often did not cite at all. Google has confirmed the behaviour is unintended and says a fix will be pushed.

Two things are worth separating here. The model’s answers were presumably still grounded in retrieved web content. What failed was the step where the system attributes each claim back to a URL and renders it. That is a plumbing failure, not a knowledge failure — which is also why it is fixable.

Why a model upgrade can break citations at all

This is the part most marketers never see, and it explains why this class of bug will keep happening.

Citations in AI search are not a natural property of an LLM. They are bolted on. A typical pipeline retrieves a set of candidate pages, feeds them to the model with instructions to answer using only those pages and to tag each claim with a source marker, then a separate layer converts those markers into clickable links.

Every one of those steps depends on the model behaving predictably. Swap in a new model version — even a better one — and its instruction-following changes subtly. It may format source markers differently. It may decide a sentence is general knowledge and not worth tagging. It may compress its output to hit latency targets and drop the markers as low-value tokens.

The downstream link-rendering layer then finds nothing to render, and fails silently. No error, no alert to the user. Just a clean, confident, unsourced answer.

Why this is a bigger deal than a normal bug

Ordinary search bugs are visible. Rankings move, you check Google Search Console, you see it. This one is close to invisible from the publisher side.

You cannot see AI Mode citation impressions the way you see classic search impressions. Google’s reporting for AI surfaces has been thin from the start, and it does not break out AI Mode as its own clean channel. So a brand could lose weeks of attribution and see only a vague dip in direct and organic traffic with no diagnosable cause.

It also lands during a period when marketing teams are being asked to justify investment in what the industry has started calling generative engine optimisation (GEO) or answer engine optimisation (AEO) — optimising to be cited by AI answers rather than just ranked. Those programmes are usually measured by citation frequency. If the citation layer itself is broken, the measurement is measuring the bug, not the strategy.

The India angle

India is one of Google’s largest user bases by volume, and AI Mode has been rolling out aggressively in the market, including in Indian languages. For Indian publishers, D2C brands, edtech companies and SaaS firms who built their acquisition on organic search, that concentration cuts both ways.

Indian marketing teams also tend to run leaner analytics stacks — often Google Analytics plus Search Console, without a separate AI-visibility tracker, which typically costs in the range of a few thousand rupees per month upward. That makes an invisible attribution failure harder to catch here than at a US publisher with a dedicated SEO analytics budget.

If you sell into India and the US from one site, worth checking whether any traffic dip is geographically uneven. A model-level bug should hit broadly, not one market — so a lopsided drop points to something else.

How to check whether you were affected

• Pull a four-to-six week trend of organic sessions and Search Console clicks side by side. A clicks drop with flat impressions and flat average position is the signature of an interface change, not a ranking change.
• Watch direct traffic. When attribution breaks, some users still find you and type your name in. Direct traffic rising while organic falls is a useful tell.
• Spot-check your top ten commercial queries in AI Mode manually. Note whether citations appear at all. Do it again after the fix ships and compare.
• Do not rebuild your content strategy this week. A bug with a confirmed fix is not a signal about your content.

What this means for you

Treat AI citation volume as a volatile metric, not a KPI. It depends on infrastructure you do not control and cannot see, and it can change overnight because someone shipped a model. Report it as a trend indicator, not a target.

Keep a manual baseline. Once a month, run your fifteen most valuable queries through AI Mode and log whether you were cited. Ten minutes of work gives you evidence when the analytics go quiet.

Build for the click that never happens. Increasingly, the reader gets your insight without visiting your site. Put your brand name, your specific numbers and your distinctive framing inside the sentences most likely to be quoted — so you get remembered even when you are not linked.

Diversify off the AI-answer surface. Email lists, communities, YouTube, LinkedIn and direct brand demand are not subject to someone else’s model deployment schedule.

Do not panic-diagnose. Before you conclude your rankings collapsed, rule out the boring explanation: the interface changed.

Frequently asked questions

What is the Google AI Mode citation bug?

It is a confirmed defect in which AI Mode answers generated by Google’s Gemini 3.8 Flash model often did not include links or citations to the web pages they were based on. Google has acknowledged the issue and said a fix will be released.

Did this bug affect my search rankings?

No. It affects whether source links are displayed alongside AI-generated answers, not how pages are ranked in classic search results. Your positions in the standard results are unrelated to it.

How would I know if my site lost traffic because of it?

Look for falling clicks in Google Search Console while impressions and average position stay stable. That pattern suggests the results interface changed rather than your ranking. Manually checking whether AI Mode cites you on your key queries adds confirmation.

Should I change my SEO strategy because of this?

Not because of this specific bug, which Google says it will fix. The broader lesson stands: AI answer surfaces are unstable, so keep manual records of your citation visibility and avoid depending on any single traffic channel.