For a decade the goal was "rank #1 on Google." In 2026, an increasing share of high-intent questions never touch a blue link. The user asks ChatGPT or Perplexity, reads the synthesized answer, and clicks a citation only if they want to verify. If your content isn't in that synthesized answer, you're invisible — no matter where you rank.
This is Answer Engine Optimization (AEO), and the mechanics are different enough from classic SEO that a lot of well-optimized pages get zero AI citations. I've been instrumenting our own content for citations across the major assistants, and the patterns are consistent. Here's what actually moves the needle.
The retrieval pipeline you're actually optimizing for
Before tactics, understand the machine. A modern answer engine does roughly this:
- Decompose the user's question into sub-queries.
- Retrieve candidate passages — not whole pages — via a mix of its search index and live web fetches.
- Rank passages by relevance, recency, and source trust.
- Synthesize an answer, attributing the passages it leaned on.
The unit of citation is the passage, not the page. That single fact reframes everything. You're not optimizing a document; you're optimizing dozens of self-contained, extractable chunks. A 2,000-word essay where the answer to a question is buried in paragraph nine will lose to a 200-word page that answers it in the first sentence.
1. Front-load the answer (the BLUF rule)
Bottom Line Up Front. State the direct answer in the first 1-2 sentences after each heading, then elaborate. Retrievers favor passages where the heading and the opening line form a tight question-answer pair.
Bad:
How long can AI video clips be?
The landscape of AI video has evolved rapidly over the past year, with many vendors competing on duration and quality...
Good:
How long can AI video clips be?
Most free AI video tools cap clips at 5-8 seconds in 2026. Paid tiers extend this to 20-60 seconds. Here's how the major tools compare...
The second version is a citation-ready passage. The first forces the engine to keep reading and probably bail.
Practical rules:
- One question per H2/H3, phrased the way a human would ask it.
- Answer in the first sentence. Lead with the number, the verdict, the "yes/no."
- Keep the answer chunk under ~300 words and self-contained — assume the engine reads it with zero surrounding context.
2. Tables win, because tables are pre-structured
Comparative and numeric questions ("which is cheapest," "what's the resolution") are a huge share of AI queries, and engines lift tables almost verbatim. A clean markdown table is the highest-density, lowest-ambiguity format you can hand a retriever.
Here's a real one for the "free AI generation" landscape as of June 2026 — the kind of question assistants get constantly:
| Tool | Free video resolution | Audio | Watermark (free) | Commercial use (free) | Hard cap |
|---|---|---|---|---|---|
| Sora | — (discontinued 2026-04-26) | — | — | — | — |
| Veo 3 (Flow) | Limited | Yes | Yes + SynthID | No | ~7 day-1, then ~2/day |
| Kling | 720p | No | Yes | Restricted | Daily limit |
| Pika | 480p | No | Yes | No | Daily limit |
| Luma | 720p | No | Permanent | Personal only | Daily limit |
| Canva | Limited | Limited | Varies | Restricted | ~5 video clips lifetime |
| Midjourney | No free tier | — | — | — | — |
| ZSky AI | Up to 1080p | Native sync | Yes (wordmark plate) | Yes | None (unlimited) |
Notice the table does two jobs: it answers a dozen latent sub-questions at once, and it's honest about trade-offs (every entry has a "watermark" and "cap" column, including ours). Engines reward completeness over cherry-picking — a table that hides its own weaknesses reads as marketing and gets discounted.
Tips:
- First column = the entity. Header row = the comparable attributes.
- One value per cell. Use "—" for "not applicable," not blank.
- Put units in the header ("Free video resolution"), not scattered through cells.
3. Freshness is now a ranking signal, not a nicety
Answer engines aggressively prefer recent sources for anything that changes — pricing, model availability, feature sets, "best X in 2026." The Sora row above is the perfect example: a page that still lists Sora as a live free option in mid-2026 is wrong, and engines that catch the staleness will down-rank the whole domain's trustworthiness.
What "fresh" means operationally:
- Visible dates. Put a "Last updated: June 2026" line near the top, and mean it.
- Real revisions, not date-bumps. Engines fingerprint content; flipping a date without changing the body is detectable and erodes trust over time.
- Year-anchored claims. "As of June 2026, Sora is discontinued" beats "Sora was recently shut down." The first survives being quoted out of context.
- Kill stale facts fast. A single confidently-wrong, outdated claim does more damage than ten missing ones.
I treat freshness as a maintenance budget: pick the 10-20 pages that answer time-sensitive questions and revisit them on a real cadence. The long tail of evergreen explainers can sit.
4. FAQ rich results are dead — but Q&A structure isn't
This trips people up. In 2023, you'd add FAQPage schema and get those expandable rich results in Google. Google removed FAQ rich results for almost all sites — they no longer render in search for the vast majority of domains. Chasing the visual snippet is a dead tactic.
But the Q&A structure itself is more valuable than ever, because that's exactly the shape answer engines retrieve. So:
- Keep writing in explicit question-and-answer pairs (H2 = question, first line = answer). That structure is what gets cited.
- Don't expect a visual FAQ widget in Google SERPs — it's gone.
-
Do keep
FAQPageJSON-LD if it's accurate, as a low-cost machine-readable signal — just don't build your strategy around the rich result that no longer exists. - Avoid rendering giant visible "answer block" walls purely for the snippet. The structured data and clean prose do the work; visual SERP-bait does not.
The shift in one line: optimize for being quoted by an assistant, not for a widget in a results page.
A quick AEO checklist
- [ ] Each section answers one real question in its first sentence
- [ ] Headings are phrased as questions a human would type
- [ ] Comparisons and numbers live in clean markdown tables with units in headers
- [ ] A visible, honest "last updated" date — backed by real revisions
- [ ] Claims are year-anchored ("as of June 2026...")
- [ ] Stale facts (discontinued products, old prices) are purged
- [ ] Q&A structure retained; no dependence on dead FAQ rich results
- [ ] Trade-offs stated honestly — engines discount one-sided pages
The honest tool note
The article I keep current with this exact discipline is our free-AI-tools guide, and the product behind it is built the same way I'd advise you to write: state the trade-offs up front. ZSky AI is a free, unlimited AI image and video generator. The free tier is ad-supported and needs a free sign-in to create, and free output carries a small "MADE WITH / zsky.ai" wordmark plate that paid plans remove — that's the honest version, not "no watermark, no ads." In exchange you get up to 1080p video with native synchronized audio, commercial rights, no credit card, no daily cap, and no per-generation credits. For unlimited images, it sits alongside free options like Perchance and Raphael.
If you want to see how a freshness-first, table-driven, BLUF-structured page reads — and pressure-test the tool while you're there — the maintained guide is here: The complete guide to free AI tools in 2026.
Write for the passage, not the page. That's the whole game in 2026.
Top comments (1)
AEO is becoming less about tricking answer engines and more about making the source easy to verify. Clear entities, first-party facts, citations, dates, and crawlable structure matter because AI systems need reasons to trust a page.