This article was AI-generated as part of an experimental historical-content project. The date reflects the period being analyzed rather than the date the article was originally written.
Most of this summer’s launches shared a property that does not show up in the announcements. More descriptions of companies are now produced in places the company cannot easily see: on a phone with no network connection, in a browser sidebar, inside an agent session logged into someone’s account, in a model running on another firm’s own servers.
That changes what a reputation audit has to cover. The spring lists on this site dealt with new answer engines and new ways in, such as tabs, photos and licensed sources. This one is about visibility: where descriptions now form out of view, and what can still be checked before the fall.
On the device and on someone else’s server
Apple’s Foundation Models framework lets any app use the on-device model behind Apple Intelligence, offline and, in Apple’s words, “using AI inference that is free of cost.” It reaches users with supported devices this fall. Expect summaries, comparisons and search features in apps you will never test, built on whatever product data, reviews and descriptions those apps already hold. The audit question is upstream: which databases, feeds and marketplaces describe your products, and are they current?
OpenAI’s gpt-oss models push in the same direction from another side. The smaller one runs with 16 GB of memory, and both carry the Apache 2.0 license. Agencies, analysts and competitors can now run a capable reasoning model on their own hardware, with no logs you could ever query. You cannot monitor those deployments. You can make sure the sources they will retrieve agree with each other.
In the browser and the agent
Perplexity’s Comet puts an assistant beside every page, able to answer questions about what the user is viewing. ChatGPT agent can browse, run tasks and, when the user takes over the browser, work inside logged-in sessions. Two checks follow.
First, read your key pages the way a sidebar would summarize them. Pricing, terms, product limits and support routes should be stated in plain text, not buried in images or scripts, because the summary is what the visitor will remember.
Second, decide who owns an agent mistake. If a customer’s agent places the wrong order, sends a confused message to your support team or acts on a misleading third-party page, your staff will meet it first. OpenAI itself warns that instructions hidden in web pages can manipulate agents. Customer service and security teams should agree now on how to recognize an agent-originated error and how to resolve it without blaming the customer.
At the defaults
GPT-5 became the default in ChatGPT on August 7, replacing GPT-4o and several other models for signed-in users. OpenAI says GPT-5 responses with web search are about 45% less likely to contain a factual error than GPT-4o. Rerun the questions where you logged errors before August and see which ones survived. A new model is a reason to retest, not to assume.
Google brought AI Mode to more than 180 new countries and territories in English on August 21. If customers or investors search in English outside the U.S., the UK and India, check AI Mode from those markets too.
The August spam update is still rolling out. Hold conclusions about ranking changes until Google marks it complete.
The remedies opinion in the Google search case leaves paid defaults in place, and the final judgment is still to come. Watch for device or browser deals that add an AI assistant to the default lineup. Those would be the first real shift.
On Wikipedia
English Wikipedia now has a speedy deletion route for unreviewed AI-generated articles, and the Foundation is trialing new bot detection. If anyone working for you has drafted Wikipedia text with a chatbot, stop and review what was submitted. Check your article’s history for the warning that it may contain text from a large language model.
None of this needs a large project. In 2023 I argued that your first impression may not be a blue link. By now it may not be on any screen your team can see, which makes the inputs to those screens the part worth checking.