By Donovan Rittenbach
Executive Summary
AEO is real, and it matters. But only a handful of AEO tactics are backed by verifiable data: crawler access, training sources like Wikipedia and Common Crawl, Perplexity’s live-retrieval model, and the Princeton GEO study’s findings on quotable language and citation lift. Everything else AEO agencies sell (such as invented timelines, uncontrolled case studies, and unvalidated “AEO scores”) is hypothesis dressed as certainty. This article separates what’s actually proven from what’s simply being sold as certainty.
The best strategy for a business learning about AEO right now starts with knowing what you don’t know.
What We Actually Know
Let’s start with the short list. These are things that are documented, verifiable, and not in dispute.
1. If a search bot is blocked, it cannot crawl your site.
Every major AI system has named crawlers. For example, OpenAI has OAI-SearchBot for indexing and ChatGPT-User for content fetching at query time. Perplexity has its own crawler, Anthropic has ClaudeBot, and Google has Googlebot. If your robots.txt blocks any of these, that engine cannot access your content. This is binary and verifiable in about thirty seconds. It is also the first and last thing most audits check.
2. Wikipedia is in every major model’s confirmed training data.
This is documented in published model cards and peer-reviewed papers. A reference citation on a Wikipedia article is one of the few AEO moves that has documented mechanism behind it.
3. Common Crawl is a confirmed pre-training source for GPT, Llama, and several other major models.
Common Crawl is a nonprofit that crawls the public web and releases the data. If your content is publicly accessible and not blocked, it has a path into model weights through Common Crawl.
4. Perplexity operates a live retrieval system, not a static knowledge base.
Perplexity’s architecture is publicly documented: when a user asks a question, Perplexity fetches current web content, ranks it for relevance and credibility, and synthesizes an answer from what it retrieves. This means freshness and crawlability are the primary levers on Perplexity.
5. ChatGPT uses two separate crawlers for two separate functions.
Search Engine Land documented this in 2026 through a honeypot experiment. OAI-SearchBot builds the index and ChatGPT-User fetches content at query time when ChatGPT decides to browse. Because these are two different systems, blocking one and not the other creates partial access that produces unreliable results.
6. Reddit influences ChatGPT’s answers more than it appears to.
A 2026 analysis of 144,000 ChatGPT citations found Reddit appeared in only 0.35% of visible citations, but occupied roughly 27% of ChatGPT’s internal search processing. This matters because optimizing for visible citations overlooks a significant portion of how influence actually works inside these systems.
7. The only peer-reviewed empirical study of AEO tactics is the GEO paper from Princeton, published at KDD 2024.
Researchers tested content modifications across 10,000 queries and measured actual citation probability changes with controls. They found that quotable language increased citation probability by up to 40%. Whereas expert quotes added approximately 41% and statistical claims added approximately 30%. Methods such as keyword stuffing reduced citation probability by about 9%. These are the only numbers in this field with a documented methodology that someone else could reproduce.
That is the complete list of things this paper will state as confirmed.
Everything else — every timeline, every percentage lift from an agency case study, every claim about which schema type moves which engine — is somewhere on a spectrum between reasonable hypothesis and invented precision. The rest of this paper explains why that distinction matters and what to do about it.
What the Field Is Selling That It Cannot Prove
Check any AEO agency’s website today and you will find claims like these: “239% citation lift in 90 days.” “AEO score: 73 out of 100.” “Ranked #1 in AI search.” “Schema optimization drives 40% more AI visibility.”
Every one of these claims has the same problem. In particular, the conditions under which the result was achieved are not disclosed, the methodology is not replicable, and the causation is not established. In every real engagement, multiple things change at once, and attributing a result to a single tactic requires controls that rarely anyone is applying.
Real-life Example #1
The “239% citation lift” figure, for example, comes from a wire service syndication study. In this study, an article distributed across a network of news domains produced a median 239% increase in brand citations. That is a real finding from a real distribution experiment, but it does not tell you how long the lift lasted, whether it held across different industries, whether the same article without syndication would have produced 20% of that lift on its own, or whether the citations translated into a single inquiry from a buyer. Despite the real number, the claim built around it (that syndication is a reliable AEO tactic with a predictable timeline and a measurable return) is not supported by the data.
Real-life Example #2
The “AEO score: 73 out of 100” problem is even more fundamental. A score requires a scale, and a scale requires a validated model of what actually predicts citation performance. However, no such validated model exists. HubSpot’s AEO grader weights sentiment at 40 points out of 100. That weighting is a design decision made by a tool builder, rather than something derived from controlled research. When a business sees their AEO score go from 61 to 74, they have no basis for concluding that anything improved — only that the tool’s arbitrary model changed its output.
This is not a critique of the people building these tools. With any new field, the data does not exist yet. But the problem is presenting the score as if the model has been validated when it has not.
Read more about the current myths in the AEO field, debunked, here.
The Measurement Problem
The question no AEO agency is answering with data
“Did the citation produce a lead?”
The entire field is measuring citation presence, which queries surface the brand, how often, on which engines, and with what sentiment. These are measurable things, so the field measures them.
But a buyer who sees your brand cited in a Perplexity response and decides not to reach out is not a win. A brand that is cited constantly and never generates an inquiry from AI-referred traffic has a citation presence and a revenue problem. The measurement and the outcome are not the same thing.
Traditional marketing learned this lesson the hard way with vanity metrics (such as impressions, reach, and engagement) that looked like success and produced no revenue. The AEO field is repeating that error with citation counts.
The honest measurement question
“Did a buyer who was already in motion (i.e. researching a purchase, evaluating options, asking AI systems about their problem) find the brand credible enough to reach out?”
Every citation, schema block, and piece of structured content is upstream of that question. If the answer to that question is unknown, the rest of the measurement is decoration.
Getting to that answer requires connecting AI citation behavior to actual lead events. That means tracking referrer logs from AI platforms, embedding a “how did you hear about us” question on inbound forms with AI-specific options, and asking buyers during discovery calls what AI systems told them and whether it influenced their decision. Despite not being technically difficult to do, none of this is standard practice in the field yet.
A Framework for Operating Honestly in Uncertain Territory
When a field has no ground truth, the honest response is not to invent precision. It is to map what is known, what is hypothesized, and what is genuinely unknown. Then, act accordingly.
Here is a practical framework for doing that.
Confirmed — things that are documented and verifiable.
- Crawler access
- Wikipedia as a training source
- Common Crawl as a training source
- Perplexity’s live retrieval architecture
- The GEO paper’s findings on content structure.
These are the tactics with documented mechanism. So, do these first, as they are the foundation everything else sits on.
Observable — things a human can measure directly with a protocol.
- Whether the brand is cited in a given query run across the four major engines
- What the citation says — accurate, inaccurate, or absent
- Which competitors are being cited instead
- How that changes between one observation period and the next
These are real measurements that require a human, a locked set of queries, and a consistent protocol.
Hypothesized — things that have plausible mechanism and some supporting evidence but no controlled validation
Most of the AEO tactics in circulation live here.
- Structured FAQ blocks
- Sequential heading architecture
- Wikidata entity creation
- G2 profile optimization
- Third-party earned media
These are reasonable bets based on what we know about how AI systems retrieve and rank content. They should be pursued, but tracked as hypotheses being tested, not as proven methods being executed.
Unknown — things that nobody in the field can tell you honestly
- How long any tactic takes to produce a citation change
- Whether a tactic that worked for one brand in one category will work for another
- What the actual weighting is between owned content and third-party citation in any model’s retrieval logic
- How model updates are reshaping citation behavior in real time
These questions do not have answers yet.
What Honest Practice Actually Looks Like
An honest AEO engagement starts with a baseline: an observation. Run a fixed set of ten queries across the four major engines. Then, record exactly what each engine says about the brand. During this process, tag each response: the brand is cited accurately, cited inaccurately, or not cited at all. Note which competitors appear instead.
That baseline is the only honest starting point because it tells you what is actually true about the brand’s current AI presence, rather than something a scoring model predicts.
Then, implement one tactic, rembering to document exactly what changed and when. At thirty days, run the same queries and record what changed. If something moved, note it — but also note every other thing that changed during that period, because any of those things might have caused the movement.
After enough clients and enough observation cycles, patterns emerge from real data. Tactic X correlated with citation improvement in twelve out of eighteen cases, with a median observation window of forty-seven days. That is a directional finding and is more honest than “239% citation lift in 90 days” because it comes with a sample size, a methodology, and an honest confidence label.
The only metric that validates any of it is leads. When a buyer’s discovery call reveals they asked an AI system about their problem and the brand appeared, that is a confirmed signal. When that happens consistently after a specific tactic was implemented, that is the closest thing to ground truth the field currently has.
Why This Matters More Than It Seems
A field that sells false certainty does two things. It extracts money from clients who cannot evaluate the claims and it delays the development of real knowledge.
The AEO field is at the moment where either of those paths is available. The comfortable path is to keep building dashboards with scores, keep publishing case studies with uncontrolled results, and keep billing for confidence that was never earned. That path is commercially viable for now because clients cannot yet evaluate the claims.
The harder path is to build the honest dataset that the field needs. This path means to run controlled observations across enough clients to produce directional findings. Furthermore, to connect citation behavior to lead generation in a way that makes the ROI case without inventing it. Also, to publish methodology openly enough that someone else could reproduce the result.
The harder path is also the one that compounds. An agency with real data in a field full of fabricated data will be in an entirely different league.
What This Paper Cannot Tell You
This paper cannot tell you which AEO tactic will produce results for your specific brand in your specific category on your specific timeline.
Nobody can tell you that honestly right now because the data does not exist. No one has done the research with enough controls and enough replication to make that claim responsibly.
What this paper can tell you is that knowing what you do not know is crucial. It is the only honest starting position in a field that is still figuring itself out. The practitioners who operate from that position — who build real observation protocols, track real lead events, and make claims only as far as their data supports — will be the ones whose methodology survives the next two years of model changes, field maturation, and client scrutiny.
Everyone else is going to have to explain why their AEO score did not predict anything.
Donovan Rittenbach is the founder of MyAIWebGuy and the developer of the Firehawk Methodology for AI citation strategy. He has thirty years of experience in web marketing and is the author of Prompt Virtuoso