GEO
An agent launch is a discovery event, not just a product launch
Anthropic has announced Claude Duet, an AI agent aimed at professionals and small business owners, launching as a research preview and available as a standalone app on iOS and Android as well as through Claude on the web. Per The Verge's report, it enters a field already occupied by rival agent products from other major labs.
The competitive horse race is the headline. The distribution shift is the substance.
Every time a general-purpose assistant becomes the first place a buyer types something like 'cheapest 3PL for EU fulfillment with Shopify integration' or 'low-MOQ supplier for custom packaging,' a slice of demand moves off the results page and into a conversation. The agent retrieves, compares, reconciles, and answers. Sometimes it links out. Often it does not.
That is the entire GEO problem compressed into one sentence: you are no longer optimizing for a position on a page, you are optimizing for inclusion in a synthesized answer — with or without a click.
From ranking to being quoted
Classic SEO assumes a human scans a results page, picks a link, and lands on your site. Agentic search collapses those steps. The model performs retrieval, extracts claims, resolves conflicts between sources, and writes prose. Your content becomes raw material rather than a destination.
Three consequences follow.
The unit of visibility changes. Rankings are increasingly replaced by citations, mentions, and paraphrases. A page that ranks sixth but contains the single clearest sentence on a topic can become the source an agent quotes. A page that ranks first but buries the answer under six paragraphs of preamble may contribute nothing at all.
Unattributed synthesis is the default failure mode. If your pricing bands, technical specs, and shipping terms exist only as prose scattered across a blog, the model has to infer them. Inference is precisely where inaccuracies enter — and where your brand ends up described incorrectly to a prospect you never meet.
Click-through becomes optional, not guaranteed. Traffic arriving from AI surfaces tends to be lower in volume and higher in intent, because the user has already been pre-qualified by the answer. Judge it against a different baseline than organic search, or you will misread the data.
What actually gets extracted
Atomic, self-contained claims
Write sentences that survive being lifted out of context. 'Our EU warehouse ships to 27 countries with a 2-day median transit time' travels well. 'As mentioned above, we handle this too' does not. Each H2 should own one question and answer it in the first two sentences, before any narrative warm-up.
Entity and schema hygiene
Agents need to know who is speaking. That means consistent organization naming across the site, a complete Organization schema with sameAs links to verifiable profiles, Product and Offer markup with currency and availability, Article markup with a real author and a real dateModified, and FAQ markup where you genuinely answer recurring questions. Disambiguate your brand against similarly named companies — agents merge entities that look alike, and merged entities produce merged (wrong) facts.
Tables, definitions, and comparisons
Structured formats outperform flowing prose in retrieval. A specification table, a definition block, a clear comparison matrix, and a short 'who this is for / who this is not for' section all give a model something precise to quote. Comparison pages matter more than ever, because that is literally the task an agent is performing on a buyer's behalf.
llms.txt as a low-cost bet
The llms.txt convention — a markdown file at the site root listing your most important pages with short descriptions — is not an official standard and is not a confirmed ranking or retrieval signal anywhere. Treat it as cheap insurance rather than a strategy. Publishing one costs an hour, creates a clean machine-readable entry point, and does no harm if it is ignored. Pair it with a genuinely crawlable site: server-rendered HTML, clean internal links, no critical content locked behind client-side rendering.
Freshness and versioning
Agents weight recency, and stale pages get quoted as if current. Put dates on facts that change, keep a visible changelog on pricing and policy pages, and retire or redirect pages that describe an older product line. An out-of-date price quoted confidently by an assistant is worse than no page at all.
Crawler access is a business decision
The indexing layer now has gatekeepers with names. Check your robots.txt rules for AI crawlers, and check whether your CDN, WAF, or bot-management configuration is silently blocking them — many teams discover their AI visibility problem is actually a 403 problem. Decide per content type: marketing and documentation pages usually want to be readable; gated research, customer portals, and proprietary datasets usually do not. The mistake is leaving that decision to a default setting nobody reviewed.
Measurement that survives the shift
Segment server logs by user agent to see how much retrieval traffic you actually receive. Track referral traffic from assistant and chat surfaces as its own channel, not folded into 'organic.' Watch branded search volume as a proxy for whether agents are naming you. Most usefully, run a fixed monthly prompt set — the ten or twenty questions your buyers actually ask — across the major assistants and record whether you appear, how you are described, and which competitors get named alongside you. That citation-share audit is the closest thing to a rank tracker this environment offers.
Where it breaks
Agents compress, and compression loses qualifiers. Expect to see your compliance caveats, regional restrictions, and eligibility conditions dropped from summaries. Counter this by making conditions short, bolded, and adjacent to the claim they modify rather than three sections away. Monitor for hallucinated prices and invented capabilities, and keep a correction path: a single canonical page per fact, clearly dated, that a model can retrieve when it re-checks.
Key takeaways
- Claude Duet is launching as a research preview across iOS, Android, and Claude on the web, aimed at professionals and small businesses — one more surface where a query can end in an answer instead of a click.
- The unit of GEO is the citation, not the ranking position. Clarity of individual claims beats page-level optimization.
- Structured, atomic, dated content is what gets extracted accurately; prose-heavy pages get inferred, and inference produces errors.
llms.txtis a low-cost, unproven bet — worth an hour, not a strategy.- Verify that your infrastructure is not blocking AI crawlers before blaming your content.
- Measure citation share with a fixed prompt set, and treat assistant referral traffic as a separate channel.
Next steps
- Run the prompt audit this week: twenty buyer questions, four assistants, one spreadsheet. Record presence, description accuracy, and co-mentioned competitors.
- Fix entity basics: Organization schema,
sameAslinks, consistent naming, author bios,dateModifiedon every commercial page. - Rewrite your five highest-value pages for extraction — answer first, table second, caveats adjacent, dates visible.
- Audit robots.txt and CDN bot rules; document the access policy per content type and who owns it.
- Publish an
llms.txtand a short internal note on what it is and is not. - Set a monthly review cadence so facts, prices, and citations are re-checked rather than assumed.
Sources
The Verge — Anthropic takes on Muse with Claude Duet, its new AI agent
