The engine behind AI answers keeps changing

In a single month this summer, Google changed the model behind its AI answers twice. First Gemini 3.5 Flash-Lite, then 3.7 Flash, rolled into Search's AI Mode about four weeks apart. That is worth sitting with. The system that decides whether your company gets named in an AI answer, and how it gets described, is now being swapped out on a monthly cadence. A visibility check you ran in July may already describe a machine that no longer exists.

The audience moved at the same time. In early August, OpenAI said ChatGPT had crossed one billion weekly users and shipped GPT-5.6. The new models cut factually wrong answers sharply, by up to 68 percent on the paid tier and 62 percent on the free default against the prior version. So a billion people a week are now asking a more careful model what to buy, who to hire, and which company to trust. More careful matters more than it sounds. A model tuned to avoid mistakes leans harder on sources it can actually verify, and quietly skips the ones it cannot.

Here is the finding that should reframe how a small company thinks about all of this. A study of just over fifty thousand commercial keywords found that paying for ads in Google's AI Mode did almost nothing to earn a citation. Ads showed up on nearly a third of those queries, and on more than half of the expensive ones, yet only about one in nine advertiser domains also appeared among the sources the AI actually cited. Paid placement, organic ranking, and AI citation are three separate channels.

You can buy your way onto the page. You cannot buy your way into the answer.

Citation is earned the slow way, through content a model can crawl, parse, and verify. Clear facts about what you do, structured so a machine reads them without guessing, published somewhere the engine is allowed to look. The faster the models get, and one of Google's new ones runs at roughly 350 tokens a second, the more they reward sources that are clean and quick to read, and the less patience they have for a site they have to fight to understand.

There is a second story running alongside the technical one, and it is being written in court. This summer a federal judge let Reddit's copyright case against Perplexity proceed, finding it plausible the company used scrapers to get around Reddit's access controls. Days earlier, News Corp sued the Brave browser over similar claims. The old assumption that anything on the open web is fair game for an AI to ingest is no longer safe. Which crawlers you allow, and what you open to them, is becoming a real decision with brand and legal weight, not a setting someone picked by default years ago.

None of this was built with small companies in mind, which is exactly why it matters to them. A large brand has a team to watch the monthly model swaps and a legal department to weigh crawler policy. A twelve-person company has neither, and is just as exposed. When an AI answer is the first impression a customer forms, and often the only one, being unreadable to the model is the modern version of being unlisted in the phone book.

So we have started treating AI visibility the way we treat any other operating metric at the companies we back. Not a one-time audit, but something you check on a schedule, because the thing you are measuring changes every few weeks. Make your facts machine-readable. Decide deliberately which answer engines you let in. Watch whether the answer about you changes when the model behind it does. The front door keeps getting rebuilt. The companies that keep a key to it are the ones that treat it as ongoing work, not a project they finished last year.