By April 2026, more people ask ChatGPT, Perplexity, and Gemini for answers than click through to traditional search results for an entire category of intent. If your marketing still treats the blue-link SERP as the only prize, you are optimizing for a shrinking surface.

Answer Engine Optimization — AEO — is the discipline of getting your content cited inside the paragraph-long responses those engines return. This post is what we've learned running it across a dozen SaaS sites over the last six months.

What changed

A Google search for "best project management software for remote teams" used to return ten links, a featured snippet, and a People Also Ask panel. In 2026 the same query on Google's Search Generative Experience opens with a multi-paragraph answer that cites three to five sources inline. The links below it still exist, but they get a smaller share of the click.

ChatGPT's web-connected mode, Perplexity's answer cards, and Gemini's AI Overviews all work on the same substrate: they pick a small set of sources, quote or paraphrase from each, and attribute. Being in that set is the new rank-one.

Three things changed with that shift:

Ranking is per-paragraph, not per-page. A page that is weak overall but contains one crisp paragraph answering a specific question can still get cited.
Query diversity exploded. Conversational prompts generate long-tail phrasings that keyword tools still catch up on. If you wait for the keyword to show up in Ahrefs before writing, you are writing late.
Citation behavior differs per engine. Perplexity cites 3–7 sources per answer and shows them visually. ChatGPT cites fewer but the ones it does cite get disproportionate referral traffic. Gemini cites within AI Overviews and sometimes hides the citations behind an expand-affordance.

What actually drives citations

After running tests across sixty pages on eight properties, a short list of signals appears in every analysis.

1. Direct-answer paragraphs at the top of the section

LLMs pull citations from paragraphs that answer the question being asked without preamble. A page structured with an H2 question and a 40–80 word direct answer immediately below it gets cited more than the same information buried in a narrative lead-in.

The pattern that works:

Code

## Does X work for Y?

Short direct answer (40-80 words, no hedging, one specific claim).

Then — only then — the context, the qualifications, the worked example.

Reversed: long intros about the history of the topic before answering the actual question. Those get skipped.

2. Statistics with attribution

LLMs prefer to cite sources that let them hedge. If your paragraph includes a specific number with an attributed source, the engine can wrap its response in "according to X" and feel confident. Unsourced numbers are worse than no numbers.

The right shape: "Session-based billing grew 34% YoY in 2025 (OpenView's 2026 SaaS Benchmarks report)." Both the number and the citation live in the paragraph.

3. Structured data the engine can skim

Schema.org matters again. FAQPage, HowTo, and Article schemas are consumed by the crawlers that feed LLM indexes. Pages with valid, specific schema get cited meaningfully more than pages with the same prose but no schema.

Do not overfit: generic "mainEntityOfPage" with boilerplate fields does nothing. FAQPage with real Q+A pairs pulled from your own content does.

4. Low latency and renderable HTML

Perplexity and Gemini's crawlers give up on pages that take over a few seconds to render. JavaScript-only content that never emits static HTML gets cited less. If your marketing site uses client-side-only rendering for key paragraphs, you are invisible to half the engines.

What does not drive citations

Keyword density. Repetition does not help. Paragraphs read by an LLM once are evaluated for whether they answer the question, not whether they contain the keyword eight times.
Word count for its own sake. A 2,500-word article does not outrank a 600-word article if the 600 words are tighter. The long articles that win are long because the question is complex, not because someone padded to hit a target.
Link equity alone. High-DA domains still get cited more in aggregate, but a low-DA page with a better direct-answer paragraph can out-cite them on specific queries. This is newer — two years ago link equity was closer to deterministic.

Practical workflow

A short cycle that works:

Find the questions users actually type into LLMs. Perplexity's "Related" suggestions and People Also Ask on Google are the cheapest sources. Paid tools that claim to mine LLM query logs are mostly guessing — check their methodology before relying on them.
Map each question to one H2 on your site. Not one page — one H2. A single long article can hold twenty H2s, each a separate AEO target.
Write the 40–80 word direct answer first. Then iterate the context around it. Most of our wins come from rewriting paragraphs where the answer is buried in the fifth sentence of a 300-word passage.
Add schema for FAQ, HowTo, or Article as applicable. Validate with Google's Rich Results Test — it is still the fastest sanity check even though the engines consuming the data are more varied now.
Measure citations, not rankings. Track referrer traffic from chat.openai.com, perplexity.ai, and the Gemini referral domain separately. Rankings in traditional tools under-represent the traffic mix.

For a shorter implementation path, use the Answer Engine Optimization checklist for SaaS teams and connect it to your Brand Setup so the content keeps the same audience and positioning language.

What to skip

Do not buy "AI SEO" tooling that promises to rewrite your content for LLM consumption in one click. The useful work is surgical — rewriting specific paragraphs, adding specific schemas. The tools that generate bulk rewrites are the same pattern that gets you demoted during Google's AI-spam updates. Which brings us to the next post.

Where MITPO fits

If you are writing the answer paragraphs by hand, MITPO's copywriting surface can help with first drafts tuned to your brand voice — but the edit you do afterward is what wins the citation. The product does not magically know the 40-word version of your answer, and anyone claiming their tool does is selling a listicle engine. The lift from having MITPO in the loop is the brand-voice guard and the speed to first draft, not a shortcut past the strategic work.