You've probably asked some version of this lately: with the advice changing every few months, what in AEO still works? The people in this thread asked it too, and they sound worn out.
Everyone is talking about AEO these days , this is working , that is working, can anyone really tell me what is working in AEO?
Some of that confusion traces back to the checklist making the rounds in marketing Slacks. It was written maybe ten months ago, a third of its items no longer work, and one or two now do harm. The mechanisms underneath have been rewired at least four times in two years, and the advice never caught up.
If you need the term, AEO (answer engine optimization) is the work of becoming the source AI answers cite, and the full glossary entry covers it. Let's go straight to the harder question of which advice is still true.
You'll get the graveyard first, since clearing dead tactics frees budget. Then come the six practices that survived everything and a three-question filter for next quarter's inventions, so you leave with a reusable method instead of another stale checklist.
The graveyard, with causes of death
Cancel any seeding retainer first. Community seeding ran from 2024 to August 2026, and for a while being in the threads engines loved to cite was the highest-ROI move in AEO, so an industry of Reddit-seeding services bloomed. Then the GPT-5.6 update swung ChatGPT's source selection toward official and licensed sources. By Otterly's measurement Reddit citations fell at least 73 percent, and the lane shut in about a week.
That's the risk with visibility you rent from a platform: one model release can reprice it. Taking part in communities for real still helps you as corroboration, but seeding them as a citation strategy no longer does.
The llms.txt line item goes next. The file has never shown a measured citation effect, Google's Search guidance says you don't need it, and the controlled tests agree. It has hung on from 2024 to whenever you read this because it's easy to sell and impossible to disprove quickly (full autopsy). Twenty minutes on one for your docs site as agent courtesy is fine, as long as nobody invoices you for it.
Your schema project needs trimming more than killing. Markup that honestly matches the visible page still earns its keep. Markup for content that isn't there, or exotic types chasing a selection edge, stopped being neutral as engines got better at spotting mismatch. Bring yours back to parity or drop it.
If your team generates pages at scale, that stopped paying in August 2026, when the spam update targeted exactly that pattern of hundreds of thin variant pages per intent. Top-10 URLs became 1.8 times more likely to vanish from the top 100, and fully automated pipelines were filtered hardest. Selection consolidating toward trusted sources did the rest, and ten deep pages now beat two hundred stubs on every surface at once.
Prompt-string sprawl is sneakier because it looks like diligence. You mine AI-flavored query logs and spin up a page per phrasing, but none of the stubs builds authority, so the cluster loses to the single deep page it replaced. You were always optimizing for the intent, never the string.
Put the five side by side and each one borrowed something, whether a platform's trust, a file's promise, a shortcut through markup or cheap volume. Each loan got repriced. The filter at the end of this post is built on that pattern, and it would have flagged all five.
Know exactly what AI says about your competitors.
RankControl's Recon Agent monitors competitor citations across ChatGPT, Perplexity, Claude, Gemini, Grok, and Google AI Mode. See where they show up and you don't.
One inherited checklist, audited live
Hold the checklist you were handed next to this representative twelve-item one, typical of what's still going around.
| Your checklist says | Do this in 2026 |
|---|---|
| "Add llms.txt" | Graveyard |
| "Post answers on Reddit and Quora" | Graveyard as strategy; corroboration only |
| "Add FAQ schema to every page" | Downgrade to parity; genuine FAQs only |
| "Create a page for every question variation" | Graveyard; consolidate |
| "Get listed on AI tool directories" | Keep: earned corroboration wearing 2024 clothes |
| "Write 40-60 word direct answers" | Keep: answer-first, sized correctly |
| "Optimize for featured snippets" | Keep: same muscle as citations |
| "Submit your site to ChatGPT" | Never existed; delete with prejudice |
| "Use conversational long-tail keywords" | Keep the intent, retire the sprawl execution |
| "Add author bios with credentials" | Keep: entity work |
| "Update content monthly" | Keep where facts age; calendar-driven elsewhere |
| "Track your AI mentions weekly" | Keep: it's the loop, and it grades all the others |
This specimen ends at five keeps, one rescue, five burials and one exorcism. If yours comes out with roughly half surviving contact with 2026, you're in line with most audits we see. Do this pass before you add anything new, since the hours it frees usually pay for the new work.
The six that survived everything
So where do the freed hours go? Into six practices that survived every model update of the last two years by targeting the only two stages that exist, retrieval and selection, rather than any surface's mood.
Your pages come first, with one intent each and the answer up top. Put the conclusion in the opening lines, sized to survive being quoted alone, and phrase headings as the questions buyers ask. Snippet-built answers quote openings and fan-out retrieval matches headings, and you can expect both to stick around.
Engines also have to reach you. Retrieval hygiene means admitting crawlers, server-rendering facts, and being present and ranking in the indexes engines actually read (Google's, Bing's, Brave's). Rankings became the qualifying round rather than the medal, and nobody has ever won by skipping qualifiers.
Then give them something to quote: falsifiable claims, real numbers, dated facts and, where you can, original data, since synthesis deduplicates sameness and attributes distinctness. A measured, quotable finding is still the single highest-yield content investment you can make.
Your self-description matters too. Use one canonical description of what you are, word for word, on your site, in directories and on profiles, so every system reconciling sources can file you with confidence. It's boring and free, and citation selection and Google's answer-layer organization both lean on it.
Other people have to vouch for you as well. Reviews, mentions, comparisons and community presence count when someone chose to include you, and the large-scale studies keep finding mentions outpredicting backlinks for AI visibility. Earned placements are also the only kind no enforcement wave can claw back.
Tying it together is the weekly loop: fixed buying queries, checked per engine every week, with dates. Citations churn hard between checks and mechanisms rewire quarterly, so the loop gives you trend lines to act on instead of anecdotes to argue about, whatever tooling you run it on.
If this sounds unexciting, good: the practices that last don't depend on platforms failing to notice. When a pitch thrills you, price in its half-life. When one sounds like a librarian's habits, it'll probably outlive the next three model updates the way these six outlived the last four.
The three-question filter
Sooner or later someone will pitch you something like this one, now making the rounds: "optimize your pages for AI agents with an agent-actions manifest." To decide whether it deserves a roadmap slot, put it through three questions.
Start with which stage it targets. Retrieval and selection are the whole pipeline, so a tactic that improves neither your findability nor your citability is aimed at a stage that doesn't exist. Most sold tactics fail right there, llms.txt being the canonical case. The manifest passes on a squint, since agents do read pages, so it's plausibly retrieval-adjacent.
Trust-earning tactics compound as they spread, while trust-borrowing ones die when platforms notice, so the second question is whether it survives everyone doing it. Imagine ten thousand sites adopting it; if the platform would just patch it, you're looking at timed arbitrage with a strategy price. The manifest soft-fails here: once every site ships one, the agents' operators decide whether to honor them, and that's platform grace rather than trust you earned.
The third asks whether you can measure it on your own queries within a month. A real tactic moves the citation rows for your fixed query set, and you can date the change against the rollout. An effect that's unmeasurable by design, or only visible in the seller's dashboard, is a story. The weekly loop exists partly so you always have an answer here. The manifest fails this one hard, because agent browsing isn't citation and you'll see no fixed-query citation effect within a month.
So the manifest belongs on your watchlist, off your roadmap, until an agent platform documents consumption, which is the llms.txt arc one era later. A tactic that scores three for three earns a pilot on a page cohort, with dates, and anything less can go under entertainment.
26 content formats. Published on your domain. Matched to your brand.
Guides, comparisons, listicles, case studies, and more. RankControl generates content that gets cited by ChatGPT, Perplexity, Claude, Gemini, Grok, and Google AI Mode.
Putting it to work this quarter
If you're the one who inherited that aging checklist, your quarter can run like this:
- In week one, audit against the graveyard. Cancel the seeding retainer, stop the programmatic sprawl, downgrade the schema project to parity, and move those hours to the six.
- In weeks two through four, work through the survivors in order: hygiene, then answer-first rewrites on your money pages, then the entity pass. Have the weekly loop live from day one so every change lands against a dated baseline.
- After that, hold the cadence and let the filter guard your roadmap.
The referee here rewrites the rules every quarter. So own what every version of the rules has rewarded: pages engines can find and answers they can lift, from a brand they can verify. Then measure closely enough to notice each rewrite the week it happens, not the quarter after. Teams working this way have shrugged through four model updates, while the checklists chasing each update's loophole have been rewritten four times and are about due for a fifth.
Your competitors are getting cited by AI. You're not.
Every day without citation tracking is a day your competitors pull ahead in ChatGPT, Perplexity, and Claude.
