AI search grew up this month and it left a paper trail. Here's what to do this week.

Three things happened in the same news cycle, and read together they close a loop most marketing teams have left open for two years. A court decided that assembling an answer is itself a publishing act. Google's own ads research team described a search box that now behaves like a conversation. And the SEO practitioners actually doing the work pointed at the one artefact nobody reads: the server log.

Here is what changed, why it matters commercially, and the single piece of work worth doing this week.

The ruling that reframes the whole discipline

The Regional Court of Aachen ordered Google to stop spreading three versions of an AI Mode answer about a local doctor, in a decision dated 24 September. Someone searching for why the physician had left his post was served a serious criminal allegation, and follow-up queries about suspension and court proceedings produced multi-paragraph accounts of a legal aftermath that was entirely real — it just belonged to a different man, also a former university professor. The court could confirm the mix-up from its own records. Fines run to 250,000 euros per violation, and the order was issued on an emergency basis before Google was heard. Google can still challenge it.

The more consequential ruling came earlier. Munich's Regional Court in May held Google directly liable for AI Overviews that tied two publishers to fraud and subscription traps using claims that appeared in none of the linked sources. Google's defence — that users could check the sources themselves — was rejected.

Why "the sources were accurate" is no longer a defence

This is the part marketers should reread. In both cases the underlying facts were true. The failure was compositional: the system stitched correct statements into an incorrect claim about a specific person or entity.

That has a direct implication for anyone who publishes. If liability attaches to the assembly rather than the source, then the platforms have a rational incentive to narrow what they assemble from. Expect tighter retrieval, more conservative attribution, and a preference for pages whose provenance is legible to a machine — clear entity signals, unambiguous dates, unambiguous authorship, claims that survive being lifted out of context and dropped into a paragraph about someone else.

Practically: vague, unattributed, entity-mushy content becomes a liability for the engine, not just a weak asset for you. The AEO work and the risk work are now the same work.

The queries arriving at your door are not the ones you optimised for

On the demand side, Google's own ads leadership described a materially different search behaviour. Ginny Marvin and Shivani Mohan used the 28 October Ads Decoded episode to explain that queries in AI Mode are three times longer than traditional searches, that one in six searches is now multimodal, and that discovery is migrating up the funnel into brainstorming sessions that happen well before a purchase decision.

Two caveats worth stating plainly, because the source states them: Marvin's post does not specify the baseline definitions or the measurement window behind the three-times and one-in-six figures. Treat them as directional, not auditable. And the claim sits alongside a counter-trend the same team flagged — that podcasts, video reviews, and firsthand experience remain part of how people research and decide. Conversational interfaces did not replace human evidence; they front-loaded it.

For B2B, the practical read is that a three-times-longer query is a three-times-more-specific intent signal. Your keyword map probably does not contain those strings. Your FAQ section, comparison pages, and pricing explainers are far closer to the shape of what is now being asked.

Your log file is the only audit trail AI search leaves

Here is the gap. Moz's walkthrough on analysing server logs for AI search opens with the uncomfortable observation that most teams have spent two to three years pouring budget into AI search while many aren't confirming whether AI bots can reach their priority pages or give up before they finish loading.

Every request leaves a line: IP, timestamp to the second, the exact path and method, the response code, the bytes served, and the user agent — often something like `ChatGPT-User/1.0`. That is enough to identify verified bots, and find problems preventing AI crawlers from receiving your content. A 403 from a WAF rule nobody remembers writing looks exactly like a page that doesn't exist, from the crawler's side.

Given the Aachen and Munich rulings, this stops being housekeeping. If engines are going to be held to what they assemble, they will increasingly be able to demonstrate what they could and could not fetch. Your logs are the same evidence.

The move to make this week

Pull thirty days of logs for your ten highest-intent pages and answer five questions.

1. Which AI user agents actually arrived? Verified ones only — user-agent strings are trivially spoofed, so cross-check against published IP ranges. 2. What status codes did they get? A wall of 403s or 429s is your problem, not the engine's. 3. How many bytes did you serve? A response far smaller than the page's real weight suggests truncation or an early bail. 4. Did they reach the content or just the shell? Client-side rendering is still the most common cause of an AI crawler seeing an empty document. 5. Which cited pages got crawled first? Compare hit order against what you actually want quoted.

One CSV, one afternoon, one list of blockers. That is a measurable input into a discipline that has spent two years being measured almost entirely by vibes.

What to watch next quarter

Whether Google appeals Aachen, and whether other jurisdictions borrow the Munich reasoning, will determine how conservative retrieval gets. Watch also for movement on the measurement-window question behind Google's three-times figure — if that gets published, the conversational-search thesis becomes auditable rather than asserted.

Until then, the defensible position is the boring one: be fetchable, be unambiguous, and keep the receipts.