What changed in ChatGPT search behavior in August 2026, and how we check for client impact
What OpenAI documented for ChatGPT search in August 2026, what it did not, and the before-and-after process we use to see what changed for each client.
Every time ChatGPT shifts, we get the same message from clients: "our numbers moved, what did OpenAI change?" August 2026 was no exception. Our answer has two parts. First, what OpenAI actually documented. Second, how we check whether that, or something else, explains a given client's data.
Sources for the first part: OpenAI's Help Center release notes and the rewritten "Searching the web with ChatGPT" help page, read on 31 October 2026, plus a public tracker that diffs that page.
What was documented
Restaurant reservations (10 August). ChatGPT can show live reservation times from OpenTable, Resy, and Yelp in the response and complete a booking from the chat. The note describes a rollout across all plans on mobile, web, and desktop, with OpenTable available globally, Resy in the US, and Yelp in the US and Canada. For hospitality and local clients this is a new kind of answer. Listings from those three partners can now appear directly in the response.
Model updates (also in the 10 August note). Plus and Pro users got an updated GPT-5.6 Sol, and GPT-5.6 Luna was to become the default for Free and Go users, with a Think button for harder questions. Different default models can phrase and order answers differently, so a shift around that date may come from the model rather than from search.
A rewritten web search help page (about 20 August). OpenAI retitled it "Searching the web with ChatGPT" and reorganized it with sections on availability, troubleshooting, reviewing sources, and location. Availability now lists Free, Go, Plus, Pro, Business, Enterprise, and Edu, and says signed-out users can search too. A change tracker that compared versions reported the documented retrieval mechanics were unchanged: ChatGPT rewrites a question into one or more targeted queries, and uses approximate location for local results. The page also encourages people to open sources and check they support the answer.
Temporary chat personalization (27 August). Users can now personalize temporary chats with memory and custom instructions. This matters when a client says "I asked it and we were not there," since personalization changes what they see.
What we did not find
We found no release note describing a change to source selection, ranking, or citation format in ChatGPT search that month. If a client's citations moved, the documentation does not explain why. Separately, Search Engine Journal reported on an August 2026 drop in Reddit citations using Promptwatch data. We do not have the detail to describe its size or cause, so we use it as a prompt to check each client's own community-source citations.
How we check client impact
We never accept a cause from a headline. Our process:
- Pick the window. The date the client noticed, plus a prior period of the same length.
- Compare on a fixed prompt set. Same prompts, same engines, same countries. We run client programs on Promptwatch, whose daily prompt panels give us a baseline, and its A/B period comparison puts the two windows next to each other.
- Separate mentions from citations. Brand mentions can hold steady while citations change, or the reverse. They point at different fixes.
- Read the fan-outs. Query fan-outs per model show whether ChatGPT is searching differently for the same prompt.
- Look at source types. Citation Trends classifies citations by content type and source type. We check community sources first (the Reddit, YouTube, and LinkedIn reports), because those are where shifts show up fastest.
- Rule out the client's side. Agent Analytics shows IP-verified AI crawler logs, so we can see if GPTBot, OAI-SearchBot, or ChatGPT-User started hitting errors. We also check for site changes and competitor launches.
Crawler logs start on the Professional plan ($245/mo), so on Essential we raise the question and, if needed, check CDN logs by hand.
How we report it
A client report on a platform change contains four things: what was announced, what we measured, what we cannot attribute, and what we will do next. We avoid language like "ChatGPT changed its algorithm" unless a source says so. A model default change, a plan availability change, and a new answer type are different events, and a client deserves to know which one they are dealing with.
For local clients, we note that restaurant booking is one example of ChatGPT answering with partner listings. If a client sells through a partner platform, we check how they appear there as well as on their own site.
Keep a baseline
The biggest lesson from months like August is that you can only detect a change if you already had a baseline. Daily runs on a stable prompt set, with the engines and countries separated, let us say "this moved after the 10th" rather than "it feels different." Without that baseline you are guessing.
No tool can predict the next change, and none can guarantee a recommendation. What a baseline gives you is speed: you find out within days, you have evidence, and you can tell the client what is known and what is not.