June 11, 2026 8 min read

ChatGPT's first 'core update': what GPT-5.5 changed in citations

When ChatGPT made GPT-5.5 its default in May 2026, about 47% of its citations shifted in 48 hours. What the SISTRIX data shows, and what to do about it.

AEO
AI Search
Citations
ChatGPT

Abstract AI answer panel built from citation signals and trusted source cards, with some cards rising and others fading

When ChatGPT made GPT-5.5 its default model on May 22–23, 2026, the sources it cites shifted by about 47% within 48 hours — a model-driven 'core update' for AI answers, per SISTRIX. For anyone tracking which pages get cited, this is the clearest evidence yet that AI visibility is volatile and has to be measured over time, not assumed.

If you woke up to a sudden jump or drop in your ChatGPT citations in late May, the most likely cause was not your own work, and not a competitor's. It was the platform changing its default model underneath you. This post unpacks what the data actually shows, what it does not prove, and how an independent measurement habit turns a scary chart into a manageable one. It is written from the observatory's chair, not a vendor's: we run an AI Visibility Index and a directory of recommended AEO agencies, and our interest is an accurate public record, not selling you a fix.

What happened on May 22–23, 2026

SISTRIX, the German search-data company, runs a daily sample of ChatGPT answers and logs which sources each answer cites. On May 22–23 the default model behind ChatGPT moved from GPT-5 mini to GPT-5.5. Across that window, the distribution of cited sources changed by roughly 47% — against a normal weekday drift of just 1–2%. SISTRIX measured this over 38 daily samples of 100,000 German-language responses each: about 3.8 million responses and more than 100 million source mentions. They named the event a 'ChatGPT Core Update', borrowing the language Google uses for its ranking algorithm changes.

A 47% shift is not noise. It is the kind of movement that, in classic search, would have SEO teams comparing notes for weeks. The difference is the trigger: not a documented ranking-system update, but a quiet swap of the underlying language model that most users never chose and were not notified about.

Three numbers worth remembering

Citation movement: about 47% of citations changed in 48 hours, versus a 1–2% baseline on a normal day. That is a core-update-scale event, not day-to-day variance.
Fewer sources per answer: the average number of cited sources fell from 30.9 to 28.4. Each answer now leans on a slightly shorter list, so the head of that list matters more.
A localization tilt: German publishers and local service brands gained citations while several international aggregators lost ground — Indeed about −47%, Tripadvisor about −53%, YouTube about −18% in this German-language sample.

There was one loud exception to the localization story. Reddit posted the single largest absolute gain, adding roughly 7,007 citations per 10,000 responses, a 59% increase. Forum and community content kept rising even as the model otherwise favored local sources — a reminder that 'authority' inside an answer engine is not always what a publisher would predict.

Correlation, not proven cause

This is the part most coverage skips, and it is the part that matters most for honest analysis. SISTRIX is explicit that the data shows a correlation in time, not a proven cause. In their words, whether the model change alone caused the citation movement, or whether the retrieval or prompt layer was also changed in parallel, cannot be determined from the outside. The default model changed and citations moved in the same window; the internal mechanism is not observable to anyone outside OpenAI.

That caveat is not a footnote — it is the whole point. AI answer systems are stacks: a model, a retrieval layer, ranking and re-ranking, safety and prompt scaffolding. Any of those can change without announcement, and from the outside you only see the output move. This is exactly why we treat 'probabilistic, not guaranteed' as the signature position of this portal: no one can promise a fixed placement in a system whose inputs shift silently, and anyone who does is waving a red flag.

Why your citations can move without you touching anything

The practical lesson of the GPT-5.5 episode is that three different forces can move your AI visibility, and only one of them is under your control. The first is your own content and technical work — the part AEO actually addresses. The second is your competitors improving their pages. The third, and the one this event spotlights, is the platform itself: a new default model, a retrieval change, a different source mix. A citation chart that drops on May 23 may say nothing about your site and everything about a model swap.

For agencies and in-house teams, this reframes the conversation with a client or executive. The honest answer to 'why did our ChatGPT citations change?' often starts with 'let's check whether the platform changed first'. Separating platform-driven movement from work-driven movement is only possible if you were already measuring continuously, with a stable set of prompts, before the shift happened.

What to do when your ChatGPT citations change

Check for a platform event before blaming your pages. Look for a default-model change, a dated 'core update' analysis, or a coordinated shift across many sites — not just yours. If everyone in your sample moved on the same day, the cause is upstream.
Compare against a fixed prompt portfolio. Movement is only readable if you sample the same questions across the same engines on a regular cadence. A prompt portfolio you have tracked for months is what turns a one-day scare into a trend line.
Watch first-citation share and source count, not just whether you appear. With the average answer citing fewer sources (30.9 to 28.4), being in the cited set is more competitive than before. Track how often you are the first source, not only whether you are present.
Re-check the fundamentals the model still rewards. Clear topical focus, decisive and quotable answers high on the page, fresh timestamps, and consistent entity information survive most model changes better than cosmetic formatting.
Localize for non-English markets. GPT-5.5's tilt toward local publishers in the German sample is a signal: for Spanish-language or other non-English audiences, content in the user's language and consistent local entity data may carry more weight than before.
Set expectations in writing. Tell clients and stakeholders, before the next model change, that AI visibility is probabilistic and platform-dependent. The teams that look competent during a core update are the ones who said this in advance.

The measurement habit this event rewards

Everything above depends on one discipline: continuous, comparable measurement. This is the entire premise behind our public AI Visibility Index, which samples a fixed portfolio of prompts across multiple engines on a regular schedule, and behind the methodology we publish openly so the numbers can be checked. The GPT-5.5 update is, in effect, a live demonstration of why a single snapshot is worthless and a time series is everything. If you want the underlying definitions — citation rate, mention rate, prompt portfolio, share of voice — the glossary on this portal defines each one in quotable form.

It also reframes how to read citation differences between platforms. Independent studies have found ChatGPT cites brand websites far less often than Perplexity does — one analysis of 34,234 responses put ChatGPT's brand-citation rate near 0.59% against Perplexity's 13.05%. A core update on one engine does not transfer to the others; each platform has its own source logic, its own update cadence, and its own volatility. Measuring per engine, not in aggregate, is the only way to see that clearly.

What this means for agencies — and for companies choosing one

For an agency, the GPT-5.5 update is a credibility test. The agencies that handle it well are the ones that already measure continuously, explain probabilistic outcomes honestly, and never sold a guaranteed placement they cannot control. If that describes how you work, getting listed in an independent directory of recommended AEO agencies is a way to be found by companies looking for exactly that discipline — you can apply through the directory's public criteria. The directory's neutrality depends on public criteria, free inclusion and non-preferential ordering.

For a company trying to choose a provider after watching your citations swing, the takeaway is simpler: favor an agency that talks about measurement and probability rather than guarantees, and use a neutral directory to find one rather than trusting a self-description. The right partner will treat a core update as a reason to show you a trend line, not a reason to panic.

A model swap can move half your AI citations overnight. The teams that stay calm are the ones who were already measuring, and who never promised a placement they could not control.

ChatGPT's first named core update will not be its last. Default models will keep changing, retrieval layers will keep evolving, and citation charts will keep jumping for reasons no outside observer can fully see. The durable response is not to chase each shift but to build the measurement habit that lets you tell platform noise from real progress — and to keep publishing the clearest, freshest, most reusable source a model can reach for. About this portal and how we work is on the about page; the index, the directory, and the glossary are the citable assets we maintain for the sector.

Frequently asked questions

Did GPT-5.5 actually cause my citations to change?

Possibly, but it cannot be proven from outside. SISTRIX measured a roughly 47% shift in cited sources in the 48 hours around the May 22–23, 2026 default-model change — a strong correlation. Whether the model itself, the retrieval layer, or the prompt layer drove it is not observable to anyone outside OpenAI. Treat it as a likely platform event, then verify against your own tracked prompts.

How many sources does ChatGPT cite per answer now?

In the SISTRIX German-language sample, the average fell from 30.9 cited sources per response before the update to 28.4 after. The list of cited sources got shorter, which makes being near the top of it more competitive than being merely present.

Should I change my AEO strategy because of one model update?

No — a single update is a reason to measure, not to overhaul. Keep a fixed prompt portfolio, watch first-citation share per engine over time, and reinforce the fundamentals (clear focus, quotable answers, freshness, entity consistency) that survive most model changes. Overreacting to one snapshot is how teams chase noise.