Why your traffic dropped when AI search arrived

A slow decline with no obvious cause is how most publishers first meet AI search. Rankings hold, the site is healthy, nothing was migrated, and the visits still fall.

Why has my traffic dropped since AI search launched?

Because the answer now happens somewhere you cannot see. An engine reads your page, composes a reply that satisfies the question completely, and the visit that used to follow never happens. Your content did the work of answering while the engine kept the audience.

How do I tell an AI-driven drop from a ranking drop?

Compare positions against clicks over the same weeks. A ranking drop moves both together, because you fell down the page. An AI-driven drop leaves positions and impressions roughly intact while clicks fall, because your result is still listed and fewer people need to open it.

Which pages lose first?

Pages that answer one question completely. Definitions, specifications, conversion tables, opening hours and simple how-to steps are the easiest for an engine to absorb and repeat. Pages that require doing something — a calculator, a booking, a login, an original dataset — hold up far better, because an answer cannot substitute for the thing itself.

How much of my content is used to train AI models?

No vendor publishes a per-site figure, so the honest answer comes from your own request logs. Count the fetches made by training agents and, more importantly, the number of distinct URLs they touched: one popular page fetched a thousand times is still one page, while ten thousand distinct URLs is most of an archive. GPTBot, ClaudeBot, CCBot and Bytespider are the four agents doing that collection today.

Does a training crawl affect the answers people get now?

No. Training crawls feed a future model release, so nothing you publish today changes an answer today. Search-index and query-time crawlers decide whether you appear in an answer this week, which is why a blanket block often removes you from answers without preventing any training at all. The crawler reference lists which agent does which job.

Can I find out what was taken before I started measuring?

Only as far back as your logs go. Most CDNs keep request logs for days rather than years, so a publisher who starts counting now is measuring the present and has no way to reconstruct the past. Starting the count is the only way to have a history worth arguing from later.

What number actually makes the case?

Crawls per visitor returned, calculated for each engine separately. Traditional search sits near five. In one illustrative measurement, AI answer engines returned a visitor roughly every 51 to 126 crawls, and one returned nothing measurable at all. Measuring that gap turns a vague sense of decline into a figure you can act on.

What can I do once I have the number?

Three responses exist, and they lead to different places: block the crawler, charge for access, or monetize the crawl so the reading is paid for whether or not a human ever arrives.

What should I check before blaming AI search?

Three things that produce the same curve and are cheaper to fix. A migration or template change that altered internal linking. A robots or canonical change that quietly removed a section from indexing. And a reporting change — consent-mode adjustments and analytics upgrades both move measured traffic without moving real traffic. Rule those out first, because an AI-driven decline is the diagnosis you reach when the ordinary explanations do not fit.