GEO Is Two Jobs, and Your Marketing Team Can Only Do One

Every guide to generative engine optimization tells you to ship Schema.org markup and an llms.txt file. I built both for this site, a JSON-LD authority graph linking every entity and llms.txt files covering the whole blog corpus, and then I ran an adversarial audit against my own design. The audit's verdict was blunt: the citation claims were over-promised, and the engineering was still worth doing. Those two findings are the whole post.

Here is why they matter to you and not just to me. AI answer engines are becoming the front door to your business. A buyer asks ChatGPT or Perplexity or Google's AI Overviews a question, reads the synthesized answer, and either sees your name in it or does not. That traffic is small today, around 1% of total web visits by Digiday's December 2025 accounting, but it converts better than anything else you have. So the question "does the AI cite us" is turning into a revenue question. And the standard answer to it, the one your marketing team is being sold, is aimed at the wrong layer.

GEO is two layers, not one

The mistake baked into most GEO advice is treating "get cited by AI" as a single job. It is two.

The first layer is retrieval eligibility: can the engine reach your content, read it, and recognize what it is about, at all? The second layer is citation selection: given everything the engine retrieved, which sources does it quote and name? These layers fail for different reasons, they are fixed by different people, and conflating them is how companies pour effort into one while bleeding out the other.

Selection is the layer everyone argues about, because it is the layer that looks like writing. Eligibility is the layer that quietly disqualifies you before the contest starts. Keep the two apart for the rest of this post and the vendor pitches start to fall into place.

Most GEO advice sells you the wrong fix

Walk into the GEO aisle and the two products on the end cap are structured-data markup and llms.txt. Both are engineering deliverables, which is part of why the pitch feels rigorous. Both also do far less for citation than the label claims.

Start with structured data. The intuition is reasonable: if you label your content with Schema.org and JSON-LD, the machine understands it better. The evidence does not cooperate. In a controlled experiment documented by the engineering team at Evil Martians, product data placed only in JSON-LD markup was missed entirely by ChatGPT, Claude, Perplexity, Gemini, and Copilot. The models tokenize the rendered text of your page; they do not parse your markup the way a search crawler does. A separate test across dozens of mid-market domains found schema-only changes moving citation by about three percent, which is inside the noise. The honest framing, and the one my own audit landed on: schema can help an engine discover and classify a page, but it does not drive whether you get extracted and quoted.

Now llms.txt, the proposed standard for a machine-readable index of your site aimed at language models. As of mid-2025 the major answer engines do not use it. Google's Gary Illyes said so directly at a July 2025 Search Central event, telling people to use normal SEO instead. A 30-day audit of crawler logs across a thousand domains found GPTBot, ClaudeBot, and PerplexityBot making zero requests for the file. It's not that llms.txt is harmful. It's that you'd be optimizing for a reader that never shows up.

None of this means the engineering is worthless. It means the engineering being sold is aimed at selection, where it barely registers, instead of eligibility, where it decides everything. Hold that thought.

What earns the citation

Here's the part the skeptics get right, and I'm going to concede it in full before I take anything back.

The founding academic study on this, the "Generative Engine Optimization" paper from researchers at Princeton, IIT Delhi, Georgia Tech, and the Allen Institute for AI, tested nine optimization tactics across ten thousand queries. Every single tactic that moved the needle was a writing tactic. Adding quotations lifted visibility around 40 percent. Adding statistics, about 33. Citing your sources, about 28. Keyword stuffing went backwards. The paper never tested a line of structured data, because the mechanism it found was content credibility, not infrastructure.

Independent citation data points the same direction. Muck Rack analyzed more than a million links cited by ChatGPT, Claude, Gemini, and Perplexity over the back half of 2025 and found roughly 82 percent came from earned media, the third-party coverage you do not own. Ahrefs studied 75,000 brands and found that branded web mentions predict AI visibility about three times more strongly than backlinks do.

Source: Ahrefs, Top Brand Visibility Factors (75,000 brands; values averaged across ChatGPT, AI Mode, and AI Overviews), Dec 2025

Read that chart and the conclusion is hard to dodge: winning the citation is mostly about being the kind of source other people quote. Statistics, named experts, quotable claims, genuine earned coverage. That's content and public-relations work, and your marketing team is exactly the right owner for it. If a vendor tells you GEO is 80 percent good content and PR, they are not wrong. Jeremy Moser of uSERP put the number at exactly that, and the data backs him.

The skeptics push one step further, and they earn it. Optimizing for a machine you can't measure can quietly cost you the search traffic you can. That risk is real. It's one more reason the content layer should stay marketing's call, judged against the channels marketing already tracks, not something engineering owns by default.

So if selection is a content discipline, where does the engineering thesis survive? In the layer the content crowd skips right over.

The layer marketing structurally can't fix

Every number in the last section assumes one thing: that your content was retrieved in the first place. Earned media, statistics, and quotable expertise only compete if the engine could reach, read, and recognize your page. Retrieval eligibility is the precondition the selection research takes for granted. And it's pure engineering.

Three checks decide it.

Notice what these have in common. Not one of them is a writing problem. Your best content marketer cannot grant a crawler access, cannot move content out of a client-side render, and cannot wire an entity graph. These are decisions that live in front-end architecture, infrastructure configuration, and your CDN's bot policy. The Cloudflare default is the cleanest example: in July 2025 the platform made AI-crawler blocking the default for new domains and handed existing customers a toggle, a change many site owners never noticed. That's an engineering reversal a marketing team wouldn't even know to look for.

This is the narrow, defensible core of the engineering thesis. Not "structured data wins citations." It does not. The claim is smaller and sturdier: content quality decides citation once you are retrieved, and engineering decides whether you are retrieved at all. The same architecture-first instinct shows up everywhere AI touches your stack, which is the through-line in why schema is the load-bearing decision in an AI diligence system and in how the structure of a knowledge corpus determines whether AI can retrieve from it.

So who owns the machine-readable surface?

Put the two layers next to your org chart and the gap jumps out.

Citation selection belongs to marketing and PR. Retrieval eligibility belongs to engineering. Almost every company I watch hand the whole of "AI search" to marketing as a content brief, the same way they once handed it the SEO checklist. Marketing dutifully works the layer it can reach, the words, and the eligibility layer falls straight through the chart. Nobody checks the bot policy. Nobody notices the client-side render. Nobody owns the entity graph. The team that was handed the goal cannot touch the part that is blocking it, and the team that can touch it was never told it was their job.

This is the same misassignment I keep writing about in other forms. It is the procurement reflex that buys a tool before redesigning the work, and it is the training reflex that blames adoption on skills when the work was never redesigned. The pattern is always the same: assign a cross-layer problem to a single-layer team and watch the seam leak.

The fix is not a reorg. It is a decision about decision rights. Name an owner for retrieval eligibility who sits in engineering, give them the three-check audit above as a standing responsibility, and let marketing keep doing the content and earned-media work it is genuinely good at. The point is to stop pretending one team can deliver both.

The executive call: ROI without the vanity metrics

Why fund any of this when AI is still 1% of your traffic? Because of how that 1% behaves.

Microsoft's analysis of more than a thousand sites found visitors arriving from AI tools signing up at roughly eleven times the rate of search visitors. A separate thirteen-month study put the conversion rate of AI referrals near 18 percent, the highest of any channel it measured. EMARKETER projects that 31.3 percent of the US population will use generative AI search in 2026; treat that as a forecast, not a fact, but the direction is not in doubt. So your highest-converting acquisition channel is the one I keep finding managed as a line item on a content calendar.

Now the discipline part, because this is where executives get fleeced. The AI-visibility tooling market is young and the measurement is shaky. One traffic study found around 70 percent of AI referrals misattributed as "direct" in Google Analytics, so most dashboards quietly undercount the channel they are selling you. Brand citations also reshuffle constantly between queries. A vendor screenshot of "you rank #1 in ChatGPT" is close to meaningless on its own.

So do not buy GEO against a citation-rank dashboard. Buy it against the two-layer model. Fund the engineering eligibility audit first, because it is cheap, finite, and gates everything else. Keep investing in the content and earned-media work that wins selection. And instrument the channel honestly enough to know which layer your spend is moving.

That sequencing is the entire ROI argument: eligibility before selection, finite engineering before open-ended content spend. It's also the cheapest place to start, because the eligibility audit is a one-time pass, not a recurring program.

If you want to know which of the two layers your organization is failing on right now, that is a diagnosable question, not a guess. Run the AI Readiness Assessment and you will get a read on where your gaps sit. If you would rather walk through it with me, scoping the eligibility audit and the ownership split is exactly what a focused advisory session is built for, or book a 15-minute call and we will find your weakest layer together. The companies that win AI citation are not the ones that bought the most markup. They are the ones that figured out the work was two jobs, and gave each one to the team that could do it.