Four dated link changes from Google. None from the other four engines.
Google has dated four changes to how it shows links in AI answers since May. OpenAI, Anthropic, Perplexity and Microsoft dated none, and no dashboard flags any of them.
up to 2%vs47%
If you measured whether AI engines cite you in the spring and measured again this month, you do not have a before-and-after. You have two snapshots of different interfaces, and only one of the five engines has told you in writing that anything moved.
That is the practical state of citation measurement right now. Google dated four changes to how links appear in its AI answers between May and October. The other four engines dated none. No analytics dashboard annotates any of it.
What Google dated, and when
Four first-party items, each with a publication date on the page:
On 6 May 2026 Google said it was “continuing to improve how we show links in our AI Search features”, that “you’ll now also see more links directly within AI responses”, and that it was “now rolling out a new feature that highlights links from your news subscriptions in AI Mode and AI Overviews”. It added a desktop hover preview on inline links.
On 27 May the Search Central documentation changelog recorded that “The preferred sources feature is starting to roll out to AI Overviews and AI Mode.”
On 3 June Google wrote that “We’ve increased the number of inline links directly within responses and added helpful website previews to encourage people to click through”, and, looking ahead, that “we’re continuing to experiment with a range of new link designs in our AI experiences to make them more useful.”
On 15 October it said it was “launching updates to show inline links higher up in responses more frequently”, that “We’ll be testing a new display in Search AI experiences to prominently surface these links in a dedicated section at the top of a response”, and that it was “starting to roll out the prominent carousels you see for timely articles in AI Search features to additional locations and languages globally.”
Read those four together and one thing stands out: none of them carries an effective date, a rollout-complete date, a locale list or a single number. The October post mixes three different states in four paragraphs, launching, testing and starting to roll out. So even for the engine that does announce, you cannot pin the day your measurement window crossed the change.
There was also a dated model change in the same period. On 19 May Google said “Starting today, we’re upgrading Search with Gemini 3.5 Flash” as “the new default model in AI Mode for everyone globally.” That matters because it lands 13 days after the May link post, inside any sensible before window.
The other four engines dated nothing
OpenAI’s release notes are detailed and dated, 220 entries with the newest on 4 November 2026. They document model and format changes. They contain no entry about where web citations or source links appear in a search answer. The one entry that mentions a Sources icon, on 5 May, is about memory: “Tap the Sources icon below the response to view relevant saved memories, past chats, and custom instructions.”
Anthropic’s release notes carry nothing about how claude.ai displays links or citations. On Perplexity and on consumer Copilot we found no provider statement about citation display either, and Perplexity’s own changelog refused our request on the day of writing, so treat that pair as unestablished rather than as a documented absence.
This is the asymmetry that breaks most measurement write-ups. Sentences of the form “the engines changed their citation display this year” are true of Google and unevidenced for the rest. Changes on ChatGPT have been reported by observers, and some of them are well documented by third parties, but an observer sighting and a provider announcement are different objects and a measurement note should not blur them. If you are writing up a ChatGPT display change, the honest label is reported by an observer, no OpenAI statement.
Google’s own second-hand record shows why that distinction is not pedantry. When AI Mode stopped showing links in early September, a Google VP replied publicly that “This isn’t working as intended, and we’ll roll out a fix soon.” A bug that suppressed links for a few days is not a design change, and nothing in Google’s announcements would have told you it happened.
No dashboard flags a display change
Search Console has a real annotation mechanism. It is explicitly not for this. Its help page says system annotations are “Automatically generated by Search Console to show issues with data processing and reporting”. Custom annotations are your own notes, capped at 120 characters, deleted after 500 days, and, in the line that matters most here, they “won’t show up in comparison mode or 24-hour views”. Comparison mode is exactly where a before-and-after gets read.
The Data anomalies page records real discontinuities when they move a metric. It carries a logging error on the Generative AI performance report for 13 to 17 August 2026, later restored. It carries a product change that moved a reported number: “As of May 7, 2026, FAQ rich results are no longer appearing in Google Search. As a result, you will see a drop in the number of reported impressions for FAQ in the performance report.” It carries nothing for the May or October AI link changes.
There is a structural reason a link move might not show up at all. Google’s own metric definitions say that for AI Overviews, “An AI Overview occupies a single position in search results, and all links in the AI Overview are assigned that same position.” If links move higher up inside the response, that is invisible in the position column by construction. Google does not draw that conclusion; it follows from its two stated rules.
Elsewhere the picture is the same. Bing’s AI Performance help says its trends “cannot be attributed to a specific cause or event” and lists system or model updates across AI experiences as one possible cause. GA4 documents definition changes in dated release notes rather than on charts: the AI Assistant channel arrived in a note dated 13 May 2026 and says nothing about whether earlier data is reprocessed, while a later note about Source Group says explicitly that “As Source Group is populated retroactively, you can use dimensions to analyze your historical data.” One retroactive, one silent, both in the same product.
How big is a display change, next to normal churn
This is the question that decides whether re-baselining is worth it, and there is real measurement on it, all of it from vendor panels.
Sistrix sampled 100,000 German-language ChatGPT responses a day for 38 days, 3.8 million responses in total. It put a typical weekday at 1 to 2 percent citation drift, an inline layout change in early May at 2 to 2.5 percent, and the GPT-5.5 switch on 22 to 23 May at 47 percent. Its own caution is the right one to carry: “Whether the model change alone was the cause, or whether the retrieval or prompt layer was also changed in parallel, cannot be determined from the outside.”
Searchable, measuring an 8 August step on ChatGPT, did the thing most before-and-afters skip: it ran a null window first, comparing two pre-change windows against each other, and reported that without that control “the raw figure reads as a 72% reshuffle when the true effect is about 12 points”. It also stated the display-versus-behaviour problem in one sentence: “If OpenAI simply started showing more of a process it had always run, the source-side results would reproduce with zero change in behavior.”
Against that, ordinary churn is enormous. Ahrefs followed every site cited on day one across five tools for 20 days. By day two, citations were already gone for 50 percent of the AI Overviews panel, 58 percent of AI Mode, 59 percent of ChatGPT, 81 percent of Gemini and 83 percent of Copilot.
| What moved | Measured drift | Source and base |
|---|---|---|
| A typical weekday, no known change | 1 to 2% | Sistrix, 100,000 German ChatGPT responses a day |
| An inline layout change | 2 to 2.5% | Sistrix, same panel, 7 to 11 May 2026 |
| A default model switch | 47% | Sistrix, same panel, 22 to 23 May 2026 |
| One day to the next, any cause | 50 to 83% of day-one citations gone | Ahrefs, 37,500 citations across five tools, 20 days |
Two readings follow. A layout change is small enough to be buried by a week of normal variance, so a handful of prompts cannot detect one. And a model change is large enough to swamp everything else, so if a model default moved inside your window, your before-and-after is measuring the model, not the display.
What to re-baseline, and what to keep
Keep the raw records. A measurement taken on 2 May is still a true statement about what the engine showed on 2 May, and that is what makes it useful later. The thing that expires is the comparison, not the observation. We said something similar about borrowed industry averages in the post on publisher traffic: the number is only as good as the base it came with, and the base has to travel with it.
Re-baseline when a model default changed inside your window, when the interface you read the answer from changed shape, or when you cannot establish either. On ChatGPT, that last case is the normal case.
Keep the comparison when the engine is Google, the window sits between two of its dated posts, and you recorded enough to show the model menu did not move under you.
And do not try to rescue a broken comparison with a bigger sample. More prompts tighten your estimate of what the engine does now. They do not reconstruct what it did before the interface changed.
Record these six fields and the next change costs you nothing
The fix is boring, which is why it works. Each time you run a measurement, store the date and time with a timezone, the engine, the model identifier if the interface exposes one, the exact prompt strings, the geography and account state, and what the answer looked like structurally: inline links, a link block at the end, or no links.
Those six fields are what let you re-read a measurement after an interface moves, instead of discarding it. They also tell you honestly when you cannot compare: if the model field changed, you have two measurements of two systems, and the right output is two numbers rather than a trend.
This is the same discipline as the fixed query set in the manual citation check method, extended by one step. The query set makes your runs comparable with each other. The recorded context is what tells you when they stopped being comparable. And when you rank what actually moves the needle, as we did in the evidence ranking, a measurement you can still interpret six months later is worth more than a tidier one you cannot.
One number sums up why this is worth the effort. Across four announcements over six months, the amount of published effect data on Google’s link changes is zero: no click figures, no percentages, no locale lists. Whatever you learn about what those changes did, you will have to measure yourself.
Frequently asked questions
Did AI engines announce changes to how they show links?
Google did, four times since May 2026, in its own blog posts and documentation changelog. OpenAI, Anthropic, Perplexity and Microsoft published nothing about where web citations or source links appear in their consumer answers over the same period.
Does Google Search Console tell me when an AI feature changed?
Not for link display. Search Console's system annotations exist "to show issues with data processing and reporting", and its Data anomalies page carries no entry for the May or October link changes. It did annotate the FAQ rich result removal, because that moved a reported metric.
How big is a link display change compared with ordinary day-to-day variation?
In the one vendor panel that measured both, much smaller than a model change. Sistrix put a typical weekday at 1 to 2 percent citation drift, an inline layout change at 2 to 2.5 percent, and the GPT-5.5 switch at 47 percent.
Is my earlier citation measurement still a valid baseline?
As a record of what the engine showed on that date, yes. As the before half of a before-and-after, only if nothing else moved in between, and on ChatGPT you cannot establish that from any provider announcement, because there is none.
What should I record so a future change does not invalidate my measurement?
The date and time, the engine, the model identifier if the interface exposes one, the exact prompt strings, and whether the answer showed inline links, a link block or no links at all. Those are the fields that let you re-read an old measurement after the interface moves.