Profound Alternative: Tracking AI Citations vs. Engineering Them

Credit where it's due. Profound tracks brand citations across 10+ AI engines, ships prompt volume data, and benchmarks you against competitors at enterprise scale. Customers include MongoDB, Zapier, and Ramp. If you need share-of-voice reporting for a board deck, it's built for exactly that.

But independent reviews keep landing on the same gap: the platform shows you what's broken, and execution lives somewhere else. Profound's answer to "now what?" is content, meaning AI-drafted listicles, FAQs, and comparison pages.

Content is a real lever. It's also the second lever. Because content only gets cited after the fetch succeeds, and the fetch is an engineering problem.

High-intent buying queries like "best X for Y," "X vs Y pricing," and "X alternative" trigger live page fetches at answer time. ChatGPT sends ChatGPT-User. Perplexity sends Perplexity-User. Claude sends Claude-User.

None of them execute JavaScript.

Googlebot renders your React app. Agent-class AI crawlers get the raw HTML response and nothing else. If your value prop lives in client-rendered components, the AI engine cites whichever competitor shipped theirs in the initial response.

We call this the two-fetch diff: load your page in a browser, then fetch it as ChatGPT-User. If the two don't match, no dashboard will save you. You're not losing on content quality. You're losing on delivery.

Most tools blur every AI crawler into one bucket. That's how sites end up blocking the wrong bot and wondering why citations died. There are three classes, doing three different jobs.

• Agent-classChatGPT-User, Claude-User, Perplexity-UserThe live fetch at answer time, for the buying queries that matter. This is the unit of measurement.
• Search-classOAI-SearchBot, Claude-SearchBot, PerplexityBotGates whether you're a retrieval candidate at all. Bingbot sits upstream of ChatGPT Search candidacy, which surprises almost everyone.
• Training-classGPTBot, ClaudeBotComparators only. Blocking GPTBot doesn't protect your citations. Blocking ChatGPT-User kills them.

Bull AI's audit runs 51 engineering signals across these 10 canonical UAs, severity-weighted, and is not a checklist. The distinction between classes is the difference between diagnosing the actual failure and guessing.

AI crawler behavior is barely documented. Engines change fetch logic, JS handling, and retrieval pipelines without announcement. No changelogs, no release notes. Published research is thin, and what exists ages in months.

The UA taxonomy is one example. So is our schema correlation dataset, our Bingbot-to-ChatGPT-Search dependency map, and our third-party citation ownership index. Every finding traces back to a probe we ran, not a paper we cited.

Third-party studies are corroboration. The foundation is first-party, and when the engines change, our findings change with them, not six months later.

Two fixes dominate AEO checklists. The data doesn't cooperate.

Ahrefs' May 2026 study found schema markup correlated -4.6% with AI Overviews visibility. llms.txt receives roughly 0.1% of AI bot requests.

Both have a legitimate job, namely index-time entity identity, and we score them there, capped at low severity. Vendors positioning them as citation fixes are selling placebo with structured data. If your AEO invoice is mostly schema deployment, you bought the wrong layer.

During a full engagement with Freeletics, answer share on tracked high-intent fitness queries moved from 15% to 68%, ahead of Nike, Peloton, and Apple Fitness+ on those queries.

Scope that claim properly: high-intent buying queries, during the engagement. Not all queries, not forever.

Because here's the part most case studies hide: after active maintenance ended, the surface decayed. That's not a caveat, it's the finding. Answer engines re-fetch and re-rank continuously, and industry tracking puts monthly citation volatility at 40 to 60%. AEO isn't a project you finish. It's a surface you hold.

A dashboard watches the decay happen. An engineering layer counters it.

Every measurement category eventually answers to the layer that owns outcomes. When a CMO asks "we're at 4% answer share, who fixes it?", budget follows the answer to that question.

The monitoring platforms see it too, which is why they're bolting on content agents. But bolted-on content doesn't touch render parity, crawl access rules, or extraction structure. The engineering layer sits upstream of everything a content agent can produce.

Honest answer: sometimes it is. If you're an enterprise brand with a dedicated AEO team, your pages are server-rendered, and your two-fetch diff comes back clean, you have an execution machine and you need measurement. Profound is built for you.

If the diff doesn't come back clean, measure later. Fix first.

Bull AI is an Engineered AEO platform, ex-OpenAI-advised, built on a 51-signal audit taxonomy and live probes across 10 canonical crawler user agents. The methodology is living: when engines change behavior, the probes re-run and the findings update. Scores render only from persisted audit runs. We don't estimate, and unverified signals ship flagged as unverified, because a visibility platform that invents data has disqualified itself from the category.

Run the two-fetch diff on your highest-intent page. If ChatGPT-User sees a different page than your customers do, you don't have a content problem. You have a rendering problem, and it's fixable.