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.
In March I put together a spring audit: one question set across every answer surface, a research report commissioned on yourself, an agent walking your task paths, facts checked as standalone components. That list still applies. The ten weeks since have added a different kind of exposure, less about which engines exist and more about how people get into them and what the engines carry along. Here is what I would add before the summer.
Test AI Mode as a conversation
AI Mode is rolling out to everyone in the U.S. as its own Search tab, with no Labs sign-up. The new unit to audit is the session. Ask the opening question, then the follow-ups a skeptical customer or reporter would ask, and note where the cited sources change as the question narrows. Google says it will graduate features from AI Mode into core Search, so what you see there is a reasonable preview.
Photograph your own products
Since April, AI Mode can answer questions about an uploaded image, using Lens to identify the objects and then searching about them. Take photos of your packaging, your storefront, your flagship product and a competitor’s look-alike, and ask what they are and whether they are any good. A brand can now be described from a picture without anyone typing its name.
Prepare a correction path for synthetic images
ChatGPT’s native image generation, live since late March, produces convincing photorealistic images, including legible text inside them. OpenAI says its images carry C2PA metadata identifying them as GPT-4o output. That helps only if someone checks. Decide now who verifies a suspicious image, where your official image library lives, and what you would publish if a fabricated product photo or executive scene started circulating.
Map which publishers covering you have AI agreements
Le Monde’s new Perplexity deal covers cited answers with links, while its earlier OpenAI agreement covers training. Wiley’s Perplexity partnership puts purchased textbooks in front of students at subscribing institutions. List the outlets and publishers that matter in your sector, note which kind of arrangement each has, and reread their coverage of you with that in mind. Licensed coverage arrives in answers with a link that signals authority, including when it is out of date.
Review what coding agents learn about you
OpenAI’s Codex went into research preview on May 16, and Anthropic made Claude Code generally available last week. If you sell to developers, your documentation, code samples and deprecation notes now feed agents that write code in other people’s repositories. Date your deprecations, keep official SDK examples current, and retire stale samples rather than leaving them up.
Fix source facts before agents memorize them
Claude Opus 4 can work for hours and, given file access, keep memory files of key facts. An error picked up early can carry across tasks. The fix is upstream and unglamorous: get leadership, subsidiary and incident facts right on your own site and in independent sources, then make sure the neutral sentence on Wikipedia agrees, using the proper channels.
Recheck leadership pages after any change
If your chief executive, chair or general counsel changed this year, check the Knowledge Panel, Wikidata, LinkedIn, your about page and the major business databases together. Leadership is among the first facts every engine resolves, and transitions are when old and new versions sit side by side longest.
The pattern across this list is that the entry points multiplied: a tab, a photo, a licensing contract, a coding agent, a memory file. Each one is small. Together they mean a company’s description is being assembled in more places, from more starting points, than a single audit was built to cover.