OpenAI released GPT-6 Astra on September 3, 2026, and Google shipped Gemini 3.8 Flash the day before. If you report on AI visibility, the obvious question is whether last month's numbers still describe anything real.
For most teams, the September launches probably didn't break the baseline, because neither one changed the model most people get by default.The change more likely to have moved your numbers happened a month earlier and got far less attention. So the useful question is narrower than "did a model ship?" You want to know whether the default changed, for which users, and whether your tracking tool's collection path changed with it.
What actually changed in the first week of September
Start with Astra. OpenAI's launch post (September 3, 2026) describes a staged rollout: a limited set of organizations first, then Plus, Pro, Business and Enterprise users and the API over the following days. The detail that matters for measurement is in OpenAI's help center. In regular chat, Astra appears as "GPT-6 Pro" on Pro, Business and Enterprise plans, and Plus users get it in ChatGPT Work and Codex. Free and Go users stay on GPT-5.6 Luna, and paid plans still default to GPT-5.6 Sol.
Astra is something a user has to pick. It doesn't replace what someone sees when they open ChatGPT and type.
Gemini 3.8 Flash went generally available in the Gemini API on September 2. On the Search side, Search Engine Journal reported that it joined AI Mode's model menu the same day for Google AI Pro and Ultra subscribers. Free AI Mode users don't get a model choice, and nothing announced that day changed their default. SEJ also pointed out that this was Google's third Flash release in six weeks.
Two headline launches, then, and both arrived as options for paying users who go looking for them.
The default swap most dashboards absorbed: August 6
On August 6, 2026, OpenAI made GPT-5.6 Luna the default for Free and Go users and rolled out a retuned GPT-5.6 Sol to Plus and Pro users in ChatGPT's Chat experience. OpenAI says ChatGPT reaches a billion people a week, so this was the change that touched the most answers. Neowin noted that the previous default change had come in May.
One line in that announcement matters to anyone measuring through the API. OpenAI says the retuned Sol runs only in ChatGPT's Chat experience, and that the version behind the API, Work and Codex didn't change. From August 6, "GPT-5.6 Sol" meant two different models depending on where you asked.
Do default swaps actually move citations? Writesonic has rerun the same 50-prompt study across successive ChatGPT models. Between GPT-5.4 and GPT-5.5, the share of citations pointing at the brand's own site fell from 56.8% to 47.2%, and fan-out searches using the site: operator dropped from 40.5% to 12.6% (Writesonic, April 2026). Between GPT-5.5 and GPT-5.6 Sol at Medium effort, site: usage jumped back to 59.3% (Writesonic, July 2026). Each of these is a single-account, single-run study by a vendor, and Writesonic says as much. The direction has still held across four studies: each version redraws the citation map.
If your tracker pools a trailing 28-day window, your August reports, and any window that closed in early September, probably straddle that swap.
Two ways an AI visibility baseline goes stale
Contamination: one window, two models
Most trackers smooth out noise by pooling several runs. That's sound practice until a model change lands inside the window.
Take a simple illustration (these numbers are made up to show the mechanics). A weekly tracker pools four runs. Two came before a swap at a 40% mention rate and two came after at 20%. The pooled figure is 30%, which describes neither model. The confidence interval around it is calculated as if one constant rate produced all four runs, so it can look reassuringly narrow while describing a population that no longer exists.
The fix is mechanical once you know when the change happened. Split the pool at the change point, and treat the post-change estimate as directional until enough new runs build up.
Divergence: measuring a surface users left
The second failure is harder to see, and it hits the most disciplined tools. A tracker that calls a pinned model through the API gets clean, reproducible data. It also keeps measuring that model after consumers have moved on. After August 6, an API tracker requesting GPT-5.6 Sol was measuring a different Sol from the one ChatGPT users got.
How big can the surface gap get? Petra Labs, an AEO agency, ran 900 trials of the same prompt across paid ChatGPT, free ChatGPT and the API on a single day (Petra Labs, May 2026). One brand's visibility swung 32 percentage points across the three. Another showed up in 15–18% of chat answers and in none of the API answers.
Every collection method fails in its own direction, so the honest move is to know which direction yours fails in. UI scrapers and consumer panels follow the product, which means they absorb default swaps without flagging them. Profound described exactly this stance at GPT-5's launch in August 2025, explaining that it monitors ChatGPT as a product rather than as individual models. API-based tools stay internally consistent and risk drifting away from what buyers see. Neither approach is wrong, but each needs a different disclosure.
The noise floor a model change has to clear
Before you blame a model for a dip, check whether the movement is bigger than ordinary rerun noise. That noise is larger than most dashboards suggest.
Ronald Sielinski's statistical framework for generative search measurement (arXiv, revised June 2026) found that confidence intervals 5–7 percentage points wide on citation share were common for SearchGPT domains. Many apparent differences between domains sat inside that noise floor. (Sielinski works at IQRush, a vendor in this space.) A separate paper from Unusual (arXiv, 2026) found that rerunning the exact same prompt on OpenAI and Anthropic models produced brand recommendation sets with a Jaccard similarity of only 0.50–0.61. Natural rephrasings of the same buying question dropped that to 0.14–0.29.
So a four-point dip the week after a launch isn't evidence of anything yet. It has to survive repeated sampling first.
Big shifts don't always come with a launch, either. In August, Reddit's share of ChatGPT citations collapsed. Promptwatch measured an 86.4% relative drop and Petra Labs put it near 80%, but the two disagreed about the cause (Hospitality.today, September 2026). Trackers can see that something moved. None of them can see why, because nobody outside OpenAI controls the version.
How to check whether your baseline survived
These steps work whatever tool you use, including a spreadsheet of manual runs.
- Keep a dated log of default changes. Record the engine, the tier (free or paid), the surface (chat, API, AI Mode) and the date. OpenAI's release notes and help center, Google's Gemini API changelog, SEJ and Search Engine Roundtable cover most of it.
- Find out which model answered each prompt. API responses report the model that served them (the
modelfield from OpenAI and Anthropic,modelVersionfrom Gemini). That returned value is more useful than the model name your code requested, because it catches changes behind an alias. If you collect from the consumer UI, capture the model identifier the page exposes. Writesonic verified it on every conversation in its studies for exactly this reason. - Split the window at a change. Don't pool across it. Report pre- and post-change estimates separately, and label the new one directional until it has enough runs behind it.
- Hold reasoning effort constant. In Writesonic's July study, comparing Sol at High effort against the old Medium default produced a 14-point "version gain" in brand citations. With effort held constant, the gain at High was 1.7 points, which is noise on 50 prompts.
- Keep a control set of prompts you haven't worked on. Chimera Marketing recommends this for attribution. If your untouched prompts move as much as your targeted ones, the engine or the season moved, not your content.
- Re-baseline when the default moves, not on launch day. A new option in a model picker changes little. A new default changes the answers most people see.
Questions to ask your AI visibility tracking vendor
The IAB's Measuring Visibility in the AI Era (August 3, 2026) separates directional data, useful for spotting trends but not for budget decisions, from decision-grade data, and it sets out disclosure expectations for providers. Model provenance fits naturally under that framing. Four questions worth sending:
- Does each stored answer record the model that produced it, and is that the model the provider returned or the one you requested?
- Which surface do you measure: the API, logged-out web sessions, logged-in paid accounts, or a panel?
- Do your trend charts mark model changes, or does the line run straight through them?
- When a change lands mid-window, do your confidence intervals reset?
When we reviewed vendor documentation in late August 2026, we couldn't find a published policy on version pinning or re-baselining from any of the major trackers. That could change quickly, and a vendor that answers these questions well deserves credit for it.
How RankSage handles model changes
RankSage measures AI visibility through official APIs. That gives us stable, reproducible data, and it means we inherit the divergence problem described above: API answers aren't identical to what a signed-in ChatGPT user sees. We'd rather say that plainly than hide it. We sample each prompt up to three times per engine per run, report 95% Wilson confidence intervals, label single-run results as directional, and pool intervals over a trailing 28-day window.
Option A (only if per-answer model provenance is live): Every stored answer also records the model identifier the provider returned. When that identifier changes, the pooling window splits at the change, and post-change figures stay labeled directional until enough runs accumulate.
Option B (if it isn't live yet): What we don't do yet is record, for every stored answer, which model the provider returned, so a pooled interval can straddle a model change today.
FAQ
Are AI visibility trackers accurate?
They're accurate about the surface they sample and less reliable about everything else. A tracker measures one collection path (API, web UI or panel) with a real margin of error, and Sielinski's work puts that margin at several percentage points for citation share. Treat any single number as an estimate with an interval, and ask which surface it came from.
Which ChatGPT model am I using?
Check the model picker at the top of the chat. As of September 2026, Free and Go users default to GPT-5.6 Luna, paid plans default to GPT-5.6 Sol, and GPT-6 Astra appears as "GPT-6 Pro" on Pro, Business and Enterprise plans, according to OpenAI's help center.
How often does ChatGPT change its default model?
Several times a year. In 2026 the default changed in May and again on August 6. Google moves at a similar pace: Gemini 3.8 Flash was its third Flash release in six weeks.
Should I re-baseline after every model launch?
No. Re-baseline when the default changes for the users you care about, or when your tool starts calling a different model. Launches that only add an option to a model picker rarely justify it.
Do API-based trackers pick up new models automatically?
Only if the tool changes the model it requests, or requests an alias the provider repoints. A tracker pinned to a dated model keeps measuring that model, which keeps it consistent but can leave it behind what consumers actually see.
Know which model you're measuring
RankSage is in early access. We're building AI visibility measurement that tells you which surface each number came from and how much to trust it, alongside GA4, Search Console and first-party behavioral data for the same pages. If that's the kind of reporting you want, join the waitlist.