How to Keep a Platform Change Log for AI Search
Keep a dated log of changes on the platforms you measure, separate from your own work log, and check it before explaining any movement. Watch the sources vendors actually publish, as of September 2026: OpenAI's ChatGPT release notes, Anthropic's release notes for the Claude apps and API, Google's Gemini Apps release notes, Google's Search Status Dashboard for ranking updates, the Search Central and Bing Webmaster blogs, and Perplexity's changelog. Log the date, the product, what changed, which of your surfaces it touches, and a link. Then annotate your charts from it. A logged change near a movement doesn't prove the change caused it, and a movement with nothing logged isn't proof it was you, because you can't assume vendors publish every change.
Your work log records what you shipped. It can't record the week a vendor changed the default model for half your sampled surfaces, and that week will show up in your numbers whether or not anyone wrote it down. The platform change log is the other half of the timeline.
Where vendors publish changes
| Source | What it covers | Where |
|---|---|---|
| ChatGPT release notes | ChatGPT features, default models, plan changes | help.openai.com |
| Claude release notes | The Claude apps; a separate page covers the API | support.claude.com, platform.claude.com |
| Gemini Apps release notes | The Gemini app | gemini.google/release-notes |
| Google Search Status Dashboard | Dated ranking updates, such as core and spam updates | status.search.google.com |
| Google Search Central blog | Search features and Search Console changes | developers.google.com/search/blog |
| Bing Webmaster blog | Bing and Bing Webmaster Tools, including AI Performance | blogs.bing.com |
| Perplexity changelog | Perplexity product updates | perplexity.ai/changelog |
The log, with real entries
These three rows are real changes taken from the vendors' public pages, as of September 2026. The surface ids and the last column are illustrative.
date,product,surface_ids,change,source,announced,expected_effect,observed_in_series
2026-08-06,ChatGPT,S-01,"GPT-5.6 Luna to become default for Free and Go (release notes)",help.openai.com,yes,"free-tier rows sample a new model",
2026-08-31,Search Console,GSC-AI,"Generative AI performance reports rolled out to all websites (Search Central blog)",developers.google.com,yes,"first-party AI data now available",
2026-09-24,Google Search,GSC;GSC-AI,"September 2026 spam update began (Search Status Dashboard)",status.search.google.com,yes,"ranking update; watch both rows",The middle row is worth noticing. A new first-party report appearing is itself a platform change: it changes what you can measure, not what happened, and a series that starts on August 31 shouldn't be read as visibility that started then.
Reading the log against a movement
When a number moves, check the log for the same window before anything else. If a change is logged, compare the size of your move with your control set and competitors on the same surfaces; a shared move of similar size is grounds for caution in both directions, since a platform shift and a real effect can happen together. If nothing is logged, don't conclude it was you. Vendors publish what they choose to, so a step change that holds, with competitors moving at the same moment, goes in the log as 'unexplained, platform-wide (suspected)'.
Keeping it cheap
A weekly skim of the sources in the table is enough for most programmes, and only changes touching a surface you actually sample go in the log. A log of every feature launch buries the few rows that matter. If a change is ambiguous, log it with 'expected effect: unknown' rather than guessing.
FAQ
Do vendors announce every model change? — We can't tell, and it's safest to assume not. Treat the log as a lower bound on what changed.
Should competitors' site changes go in this log? — Keep them in your competitor log. This one is for the systems you measure through; mixing the two makes both harder to read.
How far back should it go? — To the start of your baseline. Earlier entries are history; later is too late, because your first comparisons would have no context.
Put this to work on your own website.
Fenn finds what your customers ask, drafts the articles and site fixes, and measures what ChatGPT, Claude, Gemini, Perplexity and Grok say about you — with every change waiting for your approval.