Why a Site That Fails Every SEO Check Still Wins in ChatGPT

This is a live teardown of three real B2B software companies audited for AI search performance across citation rate, voice share, and recommendation rate. The three companies sit in different categories. They have very different technical profiles. And the ranking order for technical compliance is almost exactly the reverse of the ranking order for purchase recommendations.

The finding is not that schema hurts you or that technical work does not matter. The finding is that technical compliance alone cannot predict which company buyers are told to purchase. Understanding that gap is the thing that changes what you actually work on.

What is the difference between being cited and being recommended in AI search?

When an LLM writes an answer, it does two things. It pulls facts and definitions from sources it finds credible. Then it names vendors for the buyer to consider. Those are different jobs, and a single page rarely does both.

A citation is when your content is used as the source of a fact or a definition. The LLM borrows your explanation, may include a footnote crediting you, and moves on. A recommendation is when your brand is named as the thing the buyer should buy. You get cited because your content was the clearest explanation available. You get recommended because the LLM has enough evidence, from your pages and from what other sites say about you, to name you as the right choice for this buyer.

Getting cited often while getting recommended rarely is the most common pattern across the companies audited. That pattern has a name. A ghost citation. The LLM uses your content as a source but then names your competitors as the purchase options, because nothing in your content made a case for buying you.

Why does finishing the technical AEO checklist not predict whether you get recommended?

The first company audited, Corp, runs in AP automation and invoice processing. It has 19 JSON-LD schema types implemented correctly, a live LLMs.txt file accessible on the root, and a technical AEO score of 100 out of 100. It checks every box that standard AEO guides recommend. Its voice share in its category is 26%, placing it seventh. It appears as the top recommendation in 6% of queries.

"You can visibly have a perfect score, check all those boxes and still not be ranking," as the workshop presenter noted. "Schema does not persuade anything. It is just a labeling system."

Schema tells an LLM how sections are structured. FAQ markup is a label that says the content below it is a question. It makes content easier to parse. It does not make the content worth recommending. An explainer page about AP automation gets quoted when an LLM needs a definition. It does not get cited as the purchase choice because nothing in the explainer made a case that Corp is the right vendor for this buyer. Corp became the textbook the LLM quotes while its competitors get named as the options to buy.

Corp's 10% ghost citation rate means roughly 1 in 10 pages that gets credited as a source does not actually deliver the attributed claim in the cited page. That rate was down from a third in August, so the team is correcting it. The fix requires two steps in order. First, dedicate one page to one specific buyer question with the question as the heading. Second, publish a figure only your company can produce, so there is something worth naming rather than just quoting.

What does get a brand recommended in AI answers?

Linear is a project development tool for engineering teams. It has zero structured data, no organization schema, no FAQ markup, not one line of JSON-LD anywhere on the site. Its technical score is 80 out of 100 because the H1 text is duplicated in the DOM. By every standard AEO metric, it is the worst performer of the three companies audited.

It is named as the top recommendation in 38% of queries. That is 32 out of 46 queries, and the lowest citation rate of the three.

Its H1 reads: "The product development system for teams and agents." That sentence names a category (product development), a buyer (teams), and a differentiator (agent). An LLM writing an answer about project management tools can lift that sentence directly without adding anything. Compare it to BambooHR's H1 at the time of the audit, which described a product transition rather than stating what the product does.

A fast test for your own homepage. Remove everything except your H1 and read it to someone outside your company. If they cannot tell you exactly what you sell and who it is for, neither can an LLM. The LLM has even less context than a new reader does.

Linear's second advantage is what its content is about. Its "Now" page is a dated changelog of what the team is building, written in first-person by the people doing the work. It does not explain project management as a category. It publishes what Linear is doing on specific dates, which makes it checkable, which makes it the kind of thing other publications quote. Other sites write about Linear because Linear publishes things only Linear can know. That earned coverage is what drives the recommendation rate. Understanding how this mechanism works is the core of getting your brand into AI answers.

"Bamboo HR bought the same position with 30,000 customers and a long head start. Linear built it by publishing things only they could publish on the record with a date." That sentence from the audit is the clearest strategic summary. Brand mass creates voice share. Proprietary, dated, checkable content creates recommendation share.

What is a ghost citation and why should you track it separately from voice share?

BambooHR is named in 86% of answers in its category across 60 queries. That is the highest voice share of the three companies and first place in its category, ahead of Rippling, Workday, and ADP. It is never named as the purchase recommendation. Zero times across 60 searches.

BambooHR has 30,000 customers and external sites have written about it for years. Competitor blog posts, comparison lists, industry publications, and review sites all mention it. The audit counted 5,877 external citations referencing BambooHR, but only 2 of those citations came from pages pointing to BambooHR directly. 465 were independent sites. 74 were competitor blog posts actively teaching LLMs that BambooHR exists, in the context of alternatives to consider. The brand mass gets it named. The lack of liftable content on its own pages means LLMs never have a sentence worth repeating that makes the case for buying it.

The second problem is mechanical. Cloudflare's bot protection is actively blocking crawler access to the BambooHR site. LLMs can read the robots.txt file and some surface-level information, but live fetches fail. That has nothing to do with privacy settings. Bot protection is doing its job too aggressively. If BambooHR published proprietary hiring benchmark data from its 30,000-company customer base, the content would be unreadable by the LLMs it needs to reach. Checking whether crawlers can access your site is a five-minute task. The Catalyst AEO audit checks all major crawlers and surfaces Cloudflare blocks as a specific finding.

BambooHR's homepage H1 says the product moved from one thing to another. Its blog covers HR practices without using any of the payroll, headcount, or hiring data it holds across 30,000 companies. Anonymized aggregate benchmarks from that data would be something no competitor could publish. None of it exists on the site.

Being named 86% of the time and recommended 0% of the time is not a voice share problem. It is a content strategy problem that requires a specific fix. The fix is not technical.

How do you find out whether you are cited or recommended?

Run three queries in the LLM your buyers use most. Pick the main question buyers in your category ask before making a purchase decision. Then look for three specific things in the answer.

First, is your brand named anywhere in the answer? That is your voice share reading. Second, is a page of yours credited as a source in a footnote or citation? That is your citation rate. Third, when the answer names vendors to consider or recommends a specific purchase, is your brand in that list? That is your recommendation rate.

Almost every company finds itself in one of two positions. It is named but not recommended, like BambooHR, which means external coverage exists but on-site content is not giving LLMs a sentence worth repeating about why to buy you. Or it is cited but not recommended, like Corp, which means your explainer content is useful as a reference but nothing on your site makes the case for your product as the purchase option. Those two problems need different fixes, and the audit tells you which one you have.

Results are not deterministic. BambooHR's voice share moved from 53% to 86% between two scans a few days apart. Some of that is a real change. Some is measurement variance from non-identical question sets. Scan weekly on a consistent question set rather than treating any single scan as a permanent ranking.

What should you change on your site based on which position you are in?

If you are cited but not recommended (the Corp situation), your explainer content is already working as a source. The problem is that your product pages do not contain a sentence that makes the purchase case. Fix the H1 on your core product or service page so that it names the category, the buyer, and a specific differentiator in one liftable sentence. Then publish one piece of content built around data only your company can produce. Client results, proprietary transaction data, aggregated benchmarks from your customer base. That creates something worth naming rather than just quoting.

If you are named but not recommended (the BambooHR situation), the coverage already exists off-site. Your pages are not contributing. The first check is mechanical. Confirm that crawlers can actually read your site. The second is content. Identify the one sentence on each key page that states what you do, who it is for, and what makes you the right choice. If that sentence does not exist in the top 30% of the page, engines are unlikely to find it. Checking your technical baseline against all major crawlers takes about 20 minutes.

If your recommendation rate is strong but your citation rate is low, like Linear, the external coverage is doing the work. Keep publishing proprietary, dated, first-person content. Other people citing it is the mechanism.

Buyer questions placed as headings in the top third of a page perform better than the same question buried in an FAQ section at the bottom. LLMs read roughly the first 30% of a page in most crawls. If the answer to a buyer question is not in that window, it may not register.

External citations from third-party publications drive recommendation rate more than anything else on your own site. The best-of lists and comparison pages in your category are the primary source of recommendation traffic for most companies. Getting your brand placed on those lists is the highest-impact external action, and it compounds over time in a way that self-published explainers do not.

The data in this audit reflects scans run on Tuesday, September 9, 2026, two days before the workshop. The companies are BambooHR and Linear, both publicly covered with accessible audit data. Numbers shift between scans, so treat them as directional, not permanent.

The takeaway

A finished technical checklist is a floor, not a ceiling. Corp proved that by scoring 100 on technical and landing sixth in its category for recommendations. Schema makes content parseable. It does not make content worth recommending.

The one action worth doing this week is running three buyer queries in ChatGPT for your category and writing down whether you are named, cited, or recommended. Those are three different answers, and most companies are strong on one and missing the other two. Once you know which position you are in, the fix is specific. Run the free audit to get the technical baseline and your current recommendation rate across ChatGPT, Claude, and Perplexity before deciding where to spend time.