Read almost any AEO job description and you will find a strategy role: own AI search visibility, drive citation share, partner with content. Read the work that actually moves citations and you will find something much less glamorous — rendering, redirects, and a text file in your site root.
That gap is why these reqs stall. Hiring managers screen for the strategy and end up with someone who cannot diagnose why the answer engines never see the page in the first place.
The answer engines are not evaluating your content. They are evaluating what survived the trip to their crawler.
The crawler deciding your visibility can’t run your website
Start here, because everything else is downstream of it. Vercel and MERJ analyzed AI crawler traffic across their network and found that none of the major AI crawlers execute JavaScript. GPTBot, ClaudeBot, PerplexityBot and the rest download JavaScript files and never run them.
They pull the file. They do not execute it. Whatever your framework was going to render on the client never exists as far as the crawler is concerned.
| Crawler | Biggest share of fetches | JavaScript fetched | Executes it? |
|---|---|---|---|
| ChatGPT | 57.70% HTML | 11.50% | No |
| Claude | 35.17% images | 23.84% | No |
| Googlebot | 31.00% HTML | 15.25% | Yes |
Google’s crawlers are the exception — they run on infrastructure built to render, which is why teams often see themselves in AI Overviews but nowhere in ChatGPT or Perplexity and cannot work out why.
The practical consequence is blunt. If your marketing site is a client-rendered single-page app, your pricing page, your comparison pages and your docs are an empty shell to most of the market. No amount of content strategy fixes that. Someone has to move those routes to server rendering or static generation, and that person needs to be comfortable in your codebase — which is why this hire sits closer to technical recruiting than to a content req.
A third of what AI crawlers fetch on your site is a 404
The same analysis turned up something even less glamorous and even more fixable. ChatGPT spends 34.82% of its fetches on 404 pages. Claude is nearly identical at 34.16%. ChatGPT burns another 14.36% of its fetches following redirects.
For scale, that is happening across 569 million ChatGPT fetches and 370 million Claude fetches in a single month.
The comparison is what makes it damning. Googlebot spends 8.22% of its fetches on 404s and 1.49% on redirects. Same web, same broken links — but Google has decades of crawl scheduling behind it, and the AI crawlers do not. They are burning roughly four times as much of their budget on your dead URLs as Google does.
Put the ChatGPT numbers together and roughly half of the attention the answer engines give your site is spent on URLs that no longer resolve or that bounce somewhere else. That is not a content problem or a brand problem. That is a stale sitemap, a migration nobody finished, and internal links pointing at paths that moved two redesigns ago.
It is also the single highest-leverage thing a new AEO hire does in their first fortnight, and it is entirely invisible in a portfolio. A candidate who has done this work will talk about log files and sitemap hygiene without being prompted. A candidate who has not will talk about prompts.
Structured data is table stakes, not a lever
Here is where you should push back on the market, because the market is overselling this one.
Ahrefs ran the controlled version of the test everyone quotes casually. They tracked 1,885 pages that added JSON-LD schema between August 2025 and March 2026 against 4,000 matched control pages. The result: AI Overview citations moved −4.6%, AI Mode +2.4%, and ChatGPT +2.2%. All three are statistically indistinguishable from zero.
That sits awkwardly next to the correlation everyone cites — 53% of AI-cited pages run schema, roughly three times the rate of pages that are not cited. Both things are true. Schema is common on cited pages because pages that get cited tend to be well-built generally, not because the markup is doing the lifting.
One caveat matters for how you brief the hire: every page in that study already had 100+ AI Overview citations before the schema went on. The authors are explicit that the finding does not extend to pages AI systems cannot parse at all, where structured data may well help a page get crawled and understood in the first place.
So keep schema. It earns its place for rich results and for making your pages legible to machines. Just do not let a candidate sell you a schema rollout as a citation strategy — if they promise you citation lift from markup alone, they have not read the data.
llms.txt, and why a candidate who leads with it is guessing
llms.txt is the proposal that you publish a curated Markdown index of your site for language models to read. It is a reasonable idea and it is the first thing a lot of candidates will pitch you, because it is concrete and easy to demo.
Google’s documentation answers it directly. On llms.txt files, Google states that maintaining them “will neither harm nor help your site’s visibility or rankings in Google Search, as Google Search ignores them.” Its AI features guidance makes the same point more broadly: you do not need to create new machine-readable files or AI text files to appear in those features.
The traffic data is more damning than the policy. Ahrefs looked at 137,210 domains and found that 97% of llms.txt files received zero requests in May 2026 — nothing fetched them at all. Among the files that did get requests, AI retrieval bots accounted for 1.1% of the traffic. SEO audit tools accounted for 21.7%. The main thing reading your llms.txt file is the software checking whether you have one.
Adoption is thin for the same reason. SE Ranking analyzed nearly 300,000 domains and found only 10.13% had the file at all — nowhere near where robots.txt and sitemaps landed.
None of which makes it wrong to publish one. It costs an hour, and other systems may use it later. But the ordering tells you something about the candidate: someone who opens with llms.txt while your pricing page is client-rendered and a third of crawler fetches are 404ing has the priority order exactly backwards.
robots.txt: block the trainer without blocking the citer
This is the mistake with the worst blast radius, and it usually happens with good intentions. Someone in legal asks to keep the site out of AI training data, an engineer adds a broad disallow, and six months later nobody can work out why the brand stopped appearing in ChatGPT.
OpenAI runs separate crawlers for separate jobs, and they are documented:
| Crawler | What OpenAI uses it for | Block it and you lose |
|---|---|---|
| GPTBot | Crawling content that may train foundation models | Training inclusion only |
| OAI-SearchBot | Surfacing websites in ChatGPT’s search features | Your presence in ChatGPT search results |
| ChatGPT-User | Visiting a page when a user asks a question | The ability to be pulled into a live answer |
Blocking GPTBot is a defensible policy choice. Blocking OAI-SearchBot is deleting yourself from the channel you are trying to win. The AEO engineer worth hiring knows the difference, can explain the equivalent split on Google’s side, and will ask who owns that decision at your company before touching the file. That intersection of policy and implementation is why these roles increasingly sit alongside marketing operations rather than under content.
How to test for this skill without asking a single question
Question lists are easy to prepare for, and we already published a set of them in the AEO/GEO hiring guide. For this particular skill, a work sample separates candidates faster.
Give them your real site and thirty minutes. Ask one thing: what would you fix first, and why?
Watch the order they work in. Someone who has done this checks how your pages render before anything else, then looks for crawl waste, then structured data, then bot policy. They will tell you which fix ships first and roughly what it costs your engineering team. Someone who opens a keyword research tool and starts talking about topic clusters is a capable traditional SEO — a real hire, just not this one, and better placed through a marketing search.
The tell is specificity about your site, not fluency about the category. Anyone can describe how AI search works. The person you want will have already found the route on your site that returns an empty shell, and will be slightly annoyed about it.
For where this role sits in the market, what the four req shapes look like, and what the job pays, the companion piece is AEO/GEO hiring in 2026. If the work you are scoping leans further into pipeline systems than site infrastructure, that is a different hire again — closer to a GTM engineer.
