AEO vs SEO vs Agentic UX: Which One Does Your Shop Actually Need?

Methodology: how we know this

Every number in this article comes from two measurement runs we ran ourselves and published at agenticux.de/report. Author of the study and this article: Stephan Lucka, founder of agenticUX.

What we measured. A measurement agent requested each shop's pages the way an AI agent would: fetch the served HTML, without executing the full browser stack, and check for five machine-readable signals: Product schema in the served HTML, a reachable or declared sitemap, a price signal readable without JavaScript execution, an unambiguous add-to-cart signal in the served HTML, and a voluntary llms.txt file.

The two runs.

• June 2026, cross-industry: 89 German online shops across multiple industries. 68 were reachable for the agent, 21 blocked it outright (23.6%).
• July 2026, fashion: 100 mid-sized German fashion shops. 90 reachable, 10 blocked.

Both cohorts were verified on July 7, 2026. The June cohort was re-measured on that date, which is why the June numbers carry a July verification date.

Two honesty rules govern the numbers. First, blocked shops are reported separately, never folded into the failure counts of individual signals. Second, not-verifiable was never counted as missing. In the June run, 22 reachable shops served product data only through JavaScript, so the schema check could not return a clean yes or no. Those 22 are reported in a separate unclear category, and only 5 shops were confirmed to have no product schema at all. When you see "60% served product schema," that is 41 confirmed positives out of 68 reachable shops, with the unclear cases counted against the shop's percentage but never called failures.

What this method cannot see. It measures the machine-readable surface. It does not measure conversion, brand strength, or whether an AI system actually recommends a shop. It tells you whether an agent could work with your shop. Whether an AI system will choose you is outside the method's scope.

The three layers, defined precisely

The terms get used interchangeably in vendor pitches. They should not be.

SEO (Search Engine Optimization) answers one question: can crawlers and search systems find, fetch, and index your pages? Its raw materials are crawl access, sitemaps, internal linking, page speed, and content that matches search intent. The failure mode at this layer is blunter than most shops assume. In our June 2026 run, 21 of 89 shops (23.6%) blocked the measurement agent outright, typically via a bot-management rule at the CDN, and returned a 403 or a challenge page instead of HTML. What good looks like here: a product URL that returns HTTP 200 with full HTML to a declared crawler user agent, plus a reachable sitemap.xml (53 of 68 reachable June shops had one, 78%). SEO is two decades old, but its role changed with AI search: it is no longer the finish line. It is the entry ticket.

AEO (Answer Engine Optimization) answers a different question: when an AI system builds an answer, can it quote you as the source? AI answers do not list ten blue results. They synthesize one answer and cite the sources that made that answer defensible. The gap between ranking and citation is visible in our data: in the June run, 22 of 68 reachable shops served product data only through JavaScript, which our measurement agent, like many AI fetchers, does not execute. Those shops render fine in a human browser and can rank in classic search, yet a fetcher reading their served HTML finds no product facts to lift. They exist for Google and are ghosts for answer engines. What good looks like here: a short direct answer above the long explanation, claims connected to evidence, consistent entity names, and Product schema in the served HTML that matches the visible content. 41 of 68 reachable June shops served that schema (60%); the per-signal notes are at agenticux.de/report.

Agentic UX answers the hardest question: can an AI agent acting for a buyer actually work with your shop? That breaks into four capabilities: the agent must be able to read your pages (parse products, prices, availability from the served HTML), trust what it reads (claims backed by evidence, policies without contradictions), compare your offer against alternatives (structured facts an agent can line up side by side), and complete the task (checkout rules, shipping conditions, and return windows that are unambiguous before the agent has to commit). The characteristic failure is a contradiction no ranking tool flags: the FAQ says 30-day returns, the checkout says 14 days for sale items, and nothing marks the item as sale. A human shrugs and buys anyway. An agent that must not guess picks the shop next door. Our worked example verity-goods walks through exactly this pattern. What good looks like here: one canonical statement per policy rule, present in the served HTML, identical on product page, FAQ, and checkout.

Side-by-side comparison

Why the order is fixed

The most expensive mistake we see is treating these as a menu. They are a dependency chain, and failures cascade downward through everything above them.

A blocked crawler kills all three layers. In the June run, 21 of 89 shops (23.6%) returned nothing usable to the measurement agent. For those shops, every euro spent on content, schema, and policy clarity is invisible. There is no partial credit: an agent that receives a bot-wall or an empty shell cannot rank you, cannot quote you, cannot buy from you. This is why access is layer zero. The July fashion run showed the same failure at a lower rate: 10 of 100 shops blocked. Better, still one in ten shops spending on marketing while the front door stays locked to AI agents acting for buyers.

Missing structured product data kills AEO and Agentic UX, but leaves SEO intact. A shop can rank on Google with human-readable pages and still be unquotable. In the June run, only 41 of 68 reachable shops (60%) served confirmed Product schema in their HTML. Another 22 served product data only via JavaScript, invisible to any fetcher that does not execute it, ours included. Those shops exist in classic search and are absent from AI answers. The July fashion run was healthier at 78% confirmed schema, which tells you fashion, a category with intense feed-driven competition, has already been forced up the stack.

Contradictory policies kill only the agentic layer, and that is where the money is. A shop can be found, quoted, even recommended, and still lose the sale at the last step because the agent cannot resolve a rule, like the returns contradiction described above. This failure mode is invisible in every SEO tool you own, and it sits directly on top of the transaction.

The practical consequence: fix in dependency order. Access first, structured data second, policy and completion clarity third. Optimizing layer three while layer zero is broken is decorating a room nobody can enter.

How to test where you fail, today

You do not need our audit to locate your first failure. Three tests, each under fifteen minutes.

Test 1: fetch your product page as an AI crawler and count the structured data.

```

curl -A "GPTBot" https://yourshop.de/your-best-selling-product | grep -c "ld+json"

```

What a pass looks like: an HTTP 200 response with at least one ld+json block containing a Product or ProductGroup object, with price and availability filled in. A minimal passing block looks like this:

```json

{

"@context": "https://schema.org",

"@type": "Product",

"name": "Merino T-Shirt Classic",

"offers": { "@type": "Offer", "price": "79.00", "priceCurrency": "EUR", "availability": "https://schema.org/InStock" }

}

```

What failure looks like: a 403, a CAPTCHA page, an HTML shell with zero ld+json blocks, or schema blocks that exist but contain an Article and no Product. In our June run this single test would have flagged a problem at just over half of the 89 shops: 48 in total, made up of 21 blocked plus 27 without confirmed product schema, 22 of those 27 unclear because product data was JavaScript-only.

Test 2: ask an answer engine a buyer question in your category.

Open ChatGPT or Perplexity and ask the question your best customer would ask, for example "best sustainable merino t-shirt under 80 euros, ships to Germany, easy returns." Repeat with three phrasings. Pass rule: you are cited in at least two of the three phrasings, and every fact attributed to you (price, material, return terms) is correct. What failure looks like: one citation out of three (noise), competitors named while you are absent, or worst, you are named with wrong facts, which means the system found fragments and guessed the rest. A wrong fact in any phrasing is a fail even if you are cited.

Test 3: answer your own return-window question using only served HTML.

View the source of your product page and your returns page (View Source, because the source is what a non-executing fetcher sees, while the rendered page is what a human sees). Try to answer: what is the return window for this exact product, and are there exceptions? What a pass looks like: one unambiguous answer, present in the served HTML, consistent across product page, FAQ, and checkout. What failure looks like: the answer exists only in a PDF, only after JavaScript renders, or in two places with two different numbers. Our worked example verity-goods walks through this failure pattern: policy contradictions that a human tolerates and an agent cannot.

What the two runs say about where the market is

The gap between the June cross-industry run and the July fashion run is itself a finding.

Access first, counted against all tested shops because blocked shops are reported separately: in June, 21 of 89 shops blocked the agent (23.6%); in July, 10 of 100.

Among the shops that were reachable:

Two readings. First, price is essentially solved everywhere (96% and 97%), because price has been feed-critical for a decade. The market fixes what platforms force it to fix. Second, llms.txt shows the widest gap between the cohorts: 34% cross-industry in June, 61% in the July fashion cohort. These are different cohorts, so the gap describes categories, and the likely driver is fashion's longer history under feed-based competition rather than adoption growth over time. Since llms.txt is voluntary and costs an afternoon, it is currently the cheapest visible differentiator in the slower categories. It signals a shop that has started thinking about machine readers at all.

Our three worked examples show what the failure classes look like in practice: cartpilot for product-evidence gaps, verity-goods for policy contradictions, and northstar for expert knowledge locked in PDFs. These are illustrative scenarios carrying an "Example, Case study" label; no client results are claimed.

Decision framework: where to spend first

Start with access (SEO layer) if Test 1 returns a block, a CAPTCHA, or an error. Nothing else you do is visible until this is fixed. This was the situation for roughly one in four shops in our June cohort.

Start with AEO if you are reachable and rank in classic search, but Test 2 shows competitors cited in AI answers while you are absent, and Test 1 shows zero or thin ld+json. Your problem is quotability: facts exist but are too unstructured, unsourced, or inconsistent to lift.

Start with Agentic UX if you pass Tests 1 and 2 but fail Test 3: you are found and cited, yet an agent could not resolve your return window, bundle rules, or shipping conditions from served HTML alone. You are losing at the final step, which is the most expensive place to lose.

Avoid buying a standalone "AI visibility" tool before running the three tests. A tool that reports your AI citations is useless if the underlying failure is a blocked crawler or a contradictory policy, and both of those are free to diagnose yourself.

Avoid treating llms.txt as the fix. It is a courtesy signal, not a substitute for served product data. 34% of reachable shops in June had one; that alone recommended nothing to anyone.

FAQ

Is AEO just SEO with a new name?

No. SEO optimizes for ranking in a results list; AEO optimizes for being quoted inside a synthesized answer. The overlap is real (both need crawl access and clean structure), but the success criterion differs. A page ranks because it matches intent. A page gets cited because its facts are verifiable, consistently stated, and connected to evidence. Plenty of pages do the first and fail the second.

Does Agentic UX replace SEO?

No, it depends on it. An agent that cannot fetch your pages cannot read, trust, compare, or complete anything. In our June 2026 measurement, 21 of 89 shops blocked the measurement agent, which made every downstream investment invisible. Agentic UX is the top layer of a stack whose bottom layer is still classic crawl access.

We rank well on Google. Are we covered?

Not necessarily. Ranking proves findability, nothing more. In the June run, 96% of reachable shops served a readable price but only 60% served confirmed product schema, and 22 shops served product data only through JavaScript that non-executing fetchers never see. The policy-contradiction failure class our worked example verity-goods illustrates is invisible to every ranking tool. Run the three self-tests above; they take under an hour combined.

What is llms.txt and do we need one?

llms.txt is a voluntary convention: a plain-text file that tells AI systems where your key content lives, complementing sitemap and schema. In our measurements, 34% of reachable cross-industry shops (June) and 61% of reachable fashion shops (July) had one. Those are different cohorts, so the gap reflects category maturity, and no growth trend can be read from it. It costs an afternoon and is worth doing, as a complement to crawl access and served product data.

How do I know if AI crawlers are blocked on my shop?

Fetch your product page with an AI crawler user agent, for example curl -A "GPTBot" https://yourshop.de/product-url, and check the response. An HTTP 200 with real HTML passes. A 403, a CAPTCHA challenge, or an empty shell fails. Also check your robots.txt and your CDN's bot-management settings; many shops block agents through a firewall rule nobody remembers setting.

Why do fashion shops score better than the cross-industry average?

The July 2026 fashion cohort beat the June cross-industry cohort on every signal: 78% vs 60% product schema, 92% vs 78% sitemap, 61% vs 34% llms.txt. The likely driver is that fashion has lived under feed-based competition (price comparison, shopping feeds) longer than most categories, so its machine-readable infrastructure matured earlier. It shows the gap is closable with existing tooling.

Can one fix improve all three layers at once?

Yes: serving complete Product schema in your HTML. It strengthens classic search results, gives answer engines liftable facts for citation, and gives agents the structured price, availability, and identity data they need to compare and act. It is the highest-leverage change for shops that are reachable but weak above the access layer: 27 of 68 reachable shops in our June run lacked confirmed product schema (40%), 22 of them because product data was JavaScript-only.

How often should we re-test?

After every release that touches product pages, policies, or your CDN configuration, and at minimum quarterly. Agent-readiness degrades silently: a new bot-management rule, a theme update that moves product data into JavaScript, or a policy edit that contradicts checkout copy can each break a layer without any visible change for human visitors. The three self-tests in this article are repeatable and free.

---

Measurements: agenticUX industry reports, June 2026 (89 cross-industry German shops) and July 2026 (100 mid-sized German fashion shops), both verified July 7, 2026. Full methodology and per-signal notes at [agenticux.de/report](https://agenticux.de/report). Not-verifiable was never counted as missing; blocked shops are reported separately.