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 the conversation about AI and company reputation has assumed one shape: a large model in a data center that has read much of the web and will tell anyone what it thinks of you. Yesterday at WWDC, Apple described a different shape, and it is worth noticing before it becomes ordinary.
Apple announced that developers will be able to call the on-device model at the core of Apple Intelligence through a new Foundation Models framework. Its examples are modest: an education app that generates a quiz from a user’s notes without cloud API costs, an outdoors app with natural language search that works offline. The framework has native Swift support, Apple says access can take as few as three lines of code, and guided generation and tool calling are built in. Developers can test it now. Users with supported devices get it this fall.
A small model, on purpose
The most useful detail came from Apple’s developer session, not the keynote. The model has about 3 billion parameters, and Apple’s engineers say plainly that it is optimized for summarization, extraction and classification, and is “not designed for world knowledge or advanced reasoning.” They also describe a content tagging adapter for tag generation, entity extraction and topic detection.
Read that through a reputation lens. This is not a model that will answer “what do you know about Company X?” from training. It will take whatever text an app gives it, through the prompt or through tools the developer defines, and condense it. The description of your company that comes out depends almost entirely on what goes in.
Where the inputs come from
Think about the apps likely to adopt this first. A travel app summarizing hotel reviews. A shopping app turning product pages into a short list of pros and cons. A finance app tagging news items by company and topic. A reading app summarizing an article someone saved about your chief executive. In each case the source material is the app’s own data: its reviews, its listings, its feeds, whatever the user has open.
That changes the monitoring question. With ChatGPT or Gemini, a communications team can at least test the system, see what it says and check the citations. Here there is no single system. There are potentially thousands of small summarizers, each fed by a different data supply, running privately on phones where nobody else sees the output. Apple built it that way deliberately, and for users that is the right call. For brands it means the summary forms downstream of data you may not know the app holds.
It also changes which errors matter. A device-scale model asked to summarize forty reviews will probably produce something fluent. Whether that summary overweights a few angry reviews from an outage last year, or treats a resolved recall as current, depends on what the app retrieved and how its developer wrote the instructions. The framework makes summaries cheap. It does not make the inputs current.
What this means for reputation work
The response is upstream. The places where your products and services are reviewed, listed and described are now raw material for local summaries you will never see. Review responses, accurate listings, dated statements on your own site that a tool call can retrieve, and clean product data in the aggregators apps license all count for more when a small model does the reading.
It also means that monitoring cloud assistants, which many teams only recently started, covers a shrinking share of the places a company gets summarized. I wondered in 2023 whether Gemini would change who summarizes you first. The answer is turning out to be: more parties than anyone can list. Nobody can audit a million private summaries. You can audit the sources they are likely to be built from.
The model doesn’t know your company. That is exactly why the data about your company matters more.