AEO Explained: How Answer Engine Optimization Actually Works, and What We Do About It at Annapurna Influence

Here is the moment that made this real for me. A client of ours, a regional service business with a decent website and eight years of steady Google rankings, asked ChatGPT to recommend providers in their city. The model returned five names. Four were competitors. The fifth did not exist. Our client, who ranked third on Google for the exact same phrase, was simply absent from the answer.

That gap between "I rank" and "the AI names me" is the entire reason AEO exists.

What AEO Is, In One Paragraph

Answer Engine Optimization is the practice of structuring, publishing, and distributing content so that generative AI systems retrieve it, quote it accurately, and recommend the brand behind it inside the generated answer itself. Some teams call the same discipline GEO, short for Generative Engine Optimization. The acronyms differ. The mechanics do not. The academic term comes from Aggarwal and colleagues at Princeton and IIT Delhi, whose KDD 2024 paper defined it as a "flexible black-box optimization framework for optimizing web content visibility" in generative engines. Black-box is the operative word. Nobody outside OpenAI or Perplexity knows the retrieval weights. What we can do is control the material those systems read.

Why This Stopped Being Optional

The volume argument first. OpenAI reported 900 million weekly active ChatGPT users in February 2026, roughly double the 400 million figure from a year earlier, and the ChatGPT app crossed one billion monthly active users in June 2026 according to Sensor Tower estimates reported by Reuters. Users send somewhere near 2.5 billion prompts a day. That is not a side channel anymore.

The traffic argument is sharper. The Pew Research Center tracked 68,879 Google searches from roughly 900 US adults and found that when an AI summary appeared, users clicked a traditional result 8% of the time versus 15% without one. Clicks on the links inside the summary itself happened in about 1% of visits. Sessions ended outright 26% of the time with a summary present, against 16% without. Ahrefs, analyzing 300,000 keywords, measured a 34.5% drop in click-through rate for the top-ranked result when an AI Overview sat above it.

So the top of the funnel is being absorbed into the answer box. Now the counterintuitive part: the clicks that survive are far more valuable. Semrush's study across 500 plus high-value topics found AI-referred visitors convert at 4.4 times the rate of standard organic. Ahrefs reported that AI search delivered 0.5% of their traffic but 12.1% of their signups. Conductor's analysis of 13,770 domains puts AI referral at about 1.08% of total sessions today, growing at triple digit rates year over year.

Fewer visitors. Dramatically better ones. That is the trade the whole discipline is built around.

 

---

 

Part One: The Seven Mechanics of How AEO Actually Works

1. Answer engines retrieve, they do not rank

Every major answer engine runs on retrieval-augmented generation, usually shortened to RAG. The pipeline described in the GEO paper has three stages: a query reformulation model rewrites your question into several sub-queries, a search layer retrieves a small set of ranked sources (often just the top five, because transformer context scales quadratically in cost), and a response model synthesizes a cited answer from those sources.

The practical consequence is brutal for traditional SEO thinking. A search engine hands you ten links and lets you choose. An answer engine reads three to six sources and writes a paragraph. There is no page two. If you are not in the retrieved set, you are not in the conversation.

2. Citations have decoupled from rankings

This is the finding that surprises most marketing teams. A March 2026 analysis of 863,000 SERPs found that only 38% of AI-cited sources rank in Google's top 10. The other 62% get cited without ranking at all.

The GEO paper's data explains why this matters so much for smaller brands. When the researchers optimized all sources simultaneously, the Cite Sources method produced a 115.1% visibility increase for websites ranked fifth in search results, while the visibility of the top-ranked site fell by an average of 30.3%. Backlink authority, the single hardest asset for a small business to accumulate, matters far less when a generative model is conditioning on text rather than counting inbound links.

3. Engines read chunks, not pages

Retrieval layers do not consume an article the way a human does. They embed passages into vectors and pull the segments that best match a query. A 2,000 word essay with one buried, brilliant answer performs worse than a tightly sectioned piece where each block stands on its own.

This is why we build every article around a canonical answer near the top, self-contained H2 and H3 sections, and structured data markup such as SoftwareApplication, FAQPage, and HowTo. We wrote up the reasoning in detail in Why Chunking Matters.

4. Some optimization tactics work, and one famous one actively hurts

Here the research is unusually clean. Aggarwal and colleagues evaluated nine methods across GEO-bench, a benchmark of 10,000 queries spanning 25 domains, averaged over five random seeds. Measured on Position-Adjusted Word Count against a 19.3 baseline:

Read that last row twice. Keyword stuffing, the oldest reflex in the SEO playbook, performed worse than doing nothing at all. On Perplexity.ai, the same method scored 21.9 against a 24.1 baseline, roughly 10% below no optimization whatsoever.

Adding credible quotations, hard statistics, and named citations is what moves the needle. On Perplexity, Statistics Addition lifted Subjective Impression by about 37% and Quotation Addition lifted Position-Adjusted Word Count by 22%.

5. Combinations beat single tactics

The researchers tested pairs of the top four methods. The best combination, Fluency Optimization paired with Statistics Addition, outperformed any single strategy by more than 5.5%. Cite Sources was relatively weak alone but averaged 31.4% improvement when combined with other methods.

Nobody optimizes one variable in the real world, and the data says you should not.

6. Every engine has a different appetite

There is no single "AI SEO." A synthesis of more than 680 million tracked citations found that only about 11% of domains are cited by both ChatGPT and Perplexity. Similarweb data covering roughly 600,000 citation events put Wikipedia at 13.15% and Reddit at 11.97% of US ChatGPT citations, more than a quarter of the total between them. Perplexity leans heavily on Reddit, LinkedIn, and G2. Gemini converts traditional search signals more directly than the others.

That divergence is exactly why we monitor across three to seven engines depending on plan tier, rather than assuming a win on one surface transfers to the rest.

7. There is a propagation delay, and it is not your agency being slow

Publishing is fast. Getting Bing's index (which feeds ChatGPT Search), Perplexity's own retrieval layer, and Google's index to crawl, ingest, and stabilize on new material is not. Expect one to six months before effects settle. This is documented AEO behavior across the industry, and any vendor promising results in two weeks is selling something other than AEO.

 

---

 

SEO vs AEO: The Honest Comparison

I get asked whether AEO replaces SEO roughly once a day. It does not. They answer different questions.

Google's AI Overviews still draw on Google's index, so search fundamentals continue to matter. We treat them as complementary layers, and we said so at length in SEO vs. AEO: Two Games, One Battle for Attention.

 

---

 

Part Two: Nine Things We Actually Do at Annapurna Influence

We are a marketing company built specifically for AEO, founded in 2026 and based in Texas. Here is the workflow, in order.

1. We start with your ground truth. Category, competitors, positioning, target buyers, plus the disambiguation details almost nobody thinks to provide: trading names, aliases, real companies that share your name, and the geography you actually serve. That last set is what stops a look-alike business from polluting your reporting.

2. We scope tightly. Each project focuses on a small set of priority prompts. We would rather track a carefully researched shortlist than three hundred loosely related phrases that make a dashboard look busy.

3. We research the prompts your buyers actually type. We combine human research, search intelligence (volume, difficulty, intent, geography), and AI-assisted analysis. We sort prompts into three buyer stages: brand-aware ("Tell me about Brand X"), solution-seeking ("plumber near me"), and problem-aware ("why is my car making that noise"). That third category is where almost nobody tests, and where you can enter the consideration set before a competitor is ever named. The full method is here.

4. We baseline before we touch anything. Every downstream change gets measured against a recorded starting point rather than guessed at.

5. We produce citation-ready long-form content. Canonical answer near the top, structured data, explicit disambiguation from adjacent products, and, following the research above, real statistics, credible quotations, and named source citations rather than keyword density.

6. We distribute where the retrieval layers already look. Content that exists and content that gets retrieved are two different things. We ship each piece to publisher partner platforms and social channels the engines already index: three platforms on Starter, five on Growth, seven on Scale.

7. We wait, honestly. One to six months of propagation, communicated up front, with monthly, biweekly, or weekly reporting depending on tier.

8. We score how you are described, not just whether you appear. Our visibility score assigns each response a weighted value of 100 divided by your position in the answer. Named first is worth 100, second 50, fifth 20, absent 0. The exponential shape mirrors the same click-through decay the GEO researchers cited when they built their position-adjusted metric. Mention rate measures awareness. The composite score folds in position, accuracy, sentiment, depth, whether a source was cited, and whether there was a call to action, then weights by question difficulty: 1.00 for brand-aware, 1.25 for narrowed-problem, 1.50 for brand-unresolved problems. Being volunteered when nobody asked for you by name is worth considerably more. Full breakdown here.

9. We verify identity before scoring. If an engine names a company that merely resembles yours, it does not count. We cross-check aliases, registered name collisions, and real service area. Low-confidence matches get logged as not mentioned rather than inflating your score, with the evidence kept in the audit trail.

What We Will Not Tell You

We do not control what any AI engine outputs, and neither does anyone else. These systems are non-deterministic: ask the same question twice, minutes apart, and the answer shifts. That is precisely why a single spot-check is worthless and why we track trends across repeated runs rather than snapshots.

We also will not pretend authority is a prerequisite. In our own controlled testing, content published on a fresh domain with no ads, PR, or link building moved an AI engine's description of a product from "does not exist" to "recommended" inside 33 days. Existing authority helps. It is not the gate.

If you want to see where you currently stand before spending anything, our free AI visibility report runs ten real prompts across ChatGPT, Claude, Gemini, and Perplexity and emails you the scored breakdown in about five minutes. No credit card, no sales call.

The buyers are already asking. The only open question is whether the machine has anything accurate to say about you when they do.

 

---

 

Sources

• Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., Deshpande, A. (2024). GEO: Generative Engine Optimization. KDD '24. https://doi.org/10.1145/3637528.3671900

• Pew Research Center (July 2025). Google Users Are Less Likely to Click on Links When an AI Summary Appears in the Results. Link

• Ahrefs (2025). AI Overviews Reduce Clicks by 34.5%. Link

• Semrush (June 2025). AI Search Traffic Study. Link

• 5W Research (June 2026). The State of AI Citations 2026, 680M tracked citations. Link

• Demandsage / Reuters (2026). ChatGPT Statistics. Link