Every buyer journey in your category starts with a definition. Before anyone compares vendors or reads pricing, they type “what is [the thing]” into ChatGPT or Perplexity. Those definitional prompts are the most common shape in AI search, and they get fanned out harder than almost anything else — the engine splits “what is X” into “X meaning,” “X vs Y,” “how X works,” “X example” and goes hunting for a clean, self-contained definition to anchor the answer. If you don’t publish one, a competitor’s glossary, a Wikipedia stub, or a third-party blog fills the slot. And whoever owns the definition tends to own the first sentence of the answer your buyer reads.

Most brands treat glossaries as SEO leftovers — a thin “/glossary” folder thrown up years ago for long-tail keywords, never touched since. In AI search that same asset is one of the highest-leverage citation surfaces you have, because a definition is exactly the kind of short, factual, extractable unit retrieval systems are built to lift.

Why definitions get cited

AI engines retrieve and score isolated passages, not whole pages, and they reward content that answers in the first breath. 44.2% of all LLM citations come from the first 30% of a page’s text. A definition page is structurally pre-optimized for that: the term is the H1, the answer is the very next thing on the page. There’s no preamble to bury the lede behind.

Definitions also match the “40-word rule” perfectly. A tight 40–60 word answer placed directly under a question-form heading is the dominant cited structure, and a good definition is naturally that length — one sentence stating what the thing is, one or two giving the boundary and an example. You’re not retrofitting your sales copy into an answer unit; the format already is one.

The third reason is hierarchy. 68.7% of AI-cited pages follow a strict H1 → H2 → H3 structure, and a glossary built one term per page (or one term per H2 on a hub) is about as clean a hierarchy as exists. Each term is a labeled, answerable unit the parser can isolate without guessing.

What to publish this week

Build a real glossary, not a keyword dump. Start with the 15–25 terms your buyers actually type before they’re ready to buy — category terms, your proprietary feature names, the acronyms your competitors hide behind. Each term gets its own treatment.

Lead every entry with a standalone definition in the first 30%: one sentence that states what the term is in plain language, written so it makes sense lifted out of the page with zero surrounding context. Restate the term inside the definition — write “Answer engine optimization is the practice of…” not “It’s the practice of…” — because the engine retrieves that sentence alone and a dangling pronoun reads as noise.

Then expand below the fold: how it works, a concrete example, how it differs from the adjacent term people confuse it with. That depth is what earns the citation over a competitor’s one-liner, and it gives the fan-out something to grab for the “how X works” and “X vs Y” sub-queries.

Add DefinedTerm and DefinedTermSet schema so the structured layer agrees with the visible text. Schema won’t earn the citation on its own, but it helps the engine resolve which entity you’re defining — and on a glossary the marginal cost is near zero.

Finally, link each definition to the money page that term supports. The glossary’s job in AI search isn’t to close — it’s to be the cited source that plants your brand and your framing inside the answer, then hand the reader a path forward. Cross-link the entries to each other too, so the cluster reads as one authoritative body on the topic rather than a scatter of orphan pages.

A definition is the cheapest answer unit you’ll ever write and one of the most-quoted. If your category’s vocabulary is being defined by someone else’s page right now, that’s the first thing to take back.

Paris Roussos has been doing SEO since 1996 (co-founded a Forbes Best of the Web–winning site back in the day) and now runs a white-label AI SEO practice for agencies and brands — flat-rate, $500–$1,500/mo per client. If your top-of-funnel traffic is leaking into ChatGPT and Perplexity and you want it back, email parisroussos@gmail.com.

Define your own terms before the machine borrows someone else’s.

Ask ChatGPT, Perplexity, or Google AI Overviews “how much does [your product] cost,” and watch what happens. The answer almost never comes from your pricing page. It comes from a G2 listing, a competitor’s comparison post, a Reddit thread from 2024, or a roundup blog that hasn’t been updated in eight months. Your own page — the single source of truth, the one you obsess over — is invisible.

This isn’t bad luck. It’s a structural mismatch between how modern pricing pages are built and how AI search retrieves answers. The fix isn’t a redesign. It’s publishing the price the way the machine can actually read it.

The mechanic — why AI engines skip your pricing page

Three things are usually going wrong at once.

One — the price isn’t in the HTML. Most modern pricing pages render through a JavaScript pricing component pulling from a config file, a billing API, or a CMS block that hydrates after page load. AI answer-fetchers like `OAI-SearchBot`, `ChatGPT-User`, and `PerplexityBot` do not execute JavaScript the way Googlebot does. They fetch the HTML and parse what’s there. If your “$49/mo” lives inside a React component that mounts on the client, the bot sees `

` and nothing else. The page returns 200 OK with no price in it.

Two — there’s no answer unit. Even on server-rendered pricing pages, the number usually sits inside a tier card with a label like “Pro” and a button that says “Start free trial.” There is no sentence anywhere on the page that says “[Product] costs $49 per month for the Pro plan, billed annually, and includes [feature list].” LLMs cite paragraphs that look like answers. They don’t reassemble visual pricing cards into prose — that’s a job they push to a third party.

Three — the page reads as undated. 65% of the content AI engines reach for is from the past year (knowledge brief #4 — 28% citation lift for pages updated inside two months), but pricing pages almost never carry a visible “Updated [Month YYYY]” line or `dateModified` schema. The model can’t tell whether your $49 is from this quarter or 2022, so it defers to a review site that does show a recent date — even if that review site is wrong.

Stack those three together and you get the pattern every operator eventually notices: third parties are answering the pricing question on your behalf, with stale or inaccurate numbers, and you have no way to correct the record.

What to publish this week

You don’t need to rebuild the page. You need to add the four things the machine is looking for.

Add a plain-text answer unit in the first 30% of the page. Right under the H1, before any pricing cards render, put one sentence: “[Product] costs $X per month for the [Tier] plan, which includes [the three things buyers care about].” Add a second sentence for the next tier. This sits inside the 44.2% of the page that produces most AI citations and gives the model something to lift verbatim. Two sentences. Above the cards. Server-rendered.

Render pricing as text, not as a JS component. If your current pricing block hydrates on the client, either (a) server-render it, (b) pre-render it for AI bot user-agents, or (c) at minimum mirror the numbers as static HTML elsewhere on the page. Test it the way the bots do: `curl -A “PerplexityBot” https://yoursite.com/pricing` and look for actual dollar amounts in the response body. If you don’t see them, neither does the model.

Show a visible “Updated” dateline plus `dateModified` schema. Put “Last updated: May 2026” near the H1 and ship corresponding `Product` + `Offer` schema with `priceSpecification` and a current `dateModified`. All three surfaces — visible text, schema, response header — should agree. Don’t bump the dateline without actually changing something, but do bump it when prices, plans, or included features genuinely move.

Publish a “what you get at $X” comparison block. A simple three-column table — tier, price, what’s included — sitting in the body of the page, in plain HTML. This matches the shape of evaluative pricing prompts (“is the $49 plan worth it”) and gets pulled the same way comparison pages get pulled. Honest trade-offs (“Pro doesn’t include single sign-on — that’s Enterprise”) read as a neutral assessment and earn citations the marketing-speak version never will.

Run the four prompts again 10–14 days after shipping: “how much does [product] cost,” “[product] pricing,” “is [product] worth the price,” “[product] vs [competitor] pricing.” If your page now shows up as a citation in at least two of the four engines, the rebuild worked. If it doesn’t, you still have a render or freshness problem — usually render.

Variant B — direct, services-first

Need this done for you? Paris Roussos runs a flat-rate AI SEO service ($500–$1,500/mo per client, white-label for agencies) covering audits, schema and entity work, AI-visibility tracking, and content engineered to be cited by LLMs. Reach him at parisroussos@gmail.com.

The pricing page is the highest-intent surface you own — letting a stale Reddit thread answer for you is the most expensive AI SEO mistake nobody flags.

Every operator I talk to wants to know if ChatGPT mentions their brand. Almost none of them ask the harder question: when it does mention you, what does it actually say? I recently watched a company celebrate a Perplexity citation that described their flagship product with a feature they killed two years ago, a price point that was never real, and a “best for enterprise teams” framing — when they sell exclusively to solo founders. They were cited. They were also wrong. And the model said all of it with total confidence.

This is the blind spot in AI SEO right now. Visibility tracking tells you whether you appear. It rarely tells you whether the description attached to your name would survive a fact-check. For a brand, a confident wrong answer is worse than no answer at all — it pre-loads a prospect with bad assumptions before they ever reach your site.

The mechanic

Here is why this happens. An LLM does not “know” your company. It assembles a description from whatever sources it retrieved or trained on, weighted toward whatever it treats as authoritative. ChatGPT pulls roughly 47.9% of its citations from Wikipedia, plus a heavy share from third-party directories and aggregators. That means the sentence a model writes about you is often a blend of a stale Crunchbase profile, a 2023 review roundup, a competitor’s comparison page, and a forum thread — none of which you control, and most of which are older than your current positioning.

The model then does what models do: it smooths. It takes those fragments and produces one fluent, authoritative-sounding paragraph. Contradictions get averaged away. Outdated facts get stated in the present tense. A single loud negative review becomes neutral-sounding “users report” framing. The output reads like settled truth because the model has no mechanism for sounding uncertain about a brand it considers itself to have “enough” data on.

The compounding problem is freshness. AI engines lean toward sources that look maintained — and a five-year-old directory listing that has never changed still looks maintained to a crawler. So your own three-month-old positioning page competes against a wall of stale third-party text that, in aggregate, outweighs you. If you have never audited what the model says, you are letting the internet’s least-current version of your company do your messaging for you.

What to do this week

1. Run the description audit, not just the citation audit. Ask each major engine — ChatGPT, Perplexity, Gemini, Google AI Overviews — three questions in a clean session: “What does [brand] do?”, “Who is [brand] for?”, and “What does [brand] cost?” Paste every answer into one doc. Read them the way a skeptical prospect would. Flag each claim that is wrong, stale, or off-positioning.

2. Trace each wrong claim back to its source. Ask the model “where did you get that?” and open the cited pages. You will usually find the culprit fast: an outdated directory profile, an old press release, a review site, or your own neglected About page. The fix lives at the source, not in a prompt.

3. Correct the sources you control first, then the ones you can claim. Update your own site — pricing page, About page, product pages — with current facts stated plainly in the opening paragraph, carrying a visible “last updated” date. Then reclaim third-party profiles: Crunchbase, G2, industry directories, your LinkedIn company page. Consistency across all of them is what an entity-resolving model rewards.

4. Publish one canonical “what we are” answer unit. A short, factual block on your site — what you do, who you serve, what it costs — written in plain declarative sentences. This is the passage you want a model to lift verbatim. Make it the easiest correct answer on the entire web to find.

None of this is glamorous. It is closer to data hygiene than to content marketing. But the brands winning AI search are not just the ones that get cited — they are the ones whose cited sentence happens to be accurate, current, and on-message, because someone made it that way on purpose.

If you’re a brand that wants to be the answer LLMs reach for (not just rank on Google), Paris Roussos has been engineering search visibility for 30 years and now runs done-for-you AI SEO. Flat-rate, no-fuss. Email parisroussos@gmail.com.

Being cited by AI search is only half the win — the other half is making sure the sentence with your name in it is one you would actually sign.

Every operator I talk to is running the same panicked spreadsheet right now: AI Overviews appear on roughly 48–50% of queries (BrightEdge, Feb 2026), organic CTR on those queries has fallen 58–61% (Ahrefs, Seer), and traffic graphs in Search Console look like a slow-motion landslide. Then they cut SEO budget.

That spreadsheet is doing the wrong math. The number that matters now isn’t sessions — it’s revenue per session. And on that line, AI traffic looks dramatically better than the organic traffic it’s replacing. If you re-run the numbers properly, “losing 58% of your clicks” is often the same as “your top channel just got more profitable.” Most operators have never built the second spreadsheet.

The mechanic: AI traffic is pre-qualified, organic is browsing

Here’s the part that breaks the doom narrative. AI search traffic converts about 4.4× better than traditional organic (cited across multiple 2025 vendor studies and confirmed in tracked client data through mid-2025). And AI-referred sessions grew 527% year-over-year through mid-2025.

Why is the conversion gap that large? Because the AI engine already did the filtering for you. Someone who types “best B2B compliance vendor for SOC 2 under $20K” into ChatGPT, reads three citations, then clicks through, is not browsing. They are roughly the equivalent of a bottom-of-funnel demo request walking in cold. The traffic loss is mostly informational, low-intent visits — the ones that bounced off your blog post anyway. What’s left is hotter.

The corollary nobody likes is this: not all CTR drops are equal. Brands cited inside AI Overviews see a +35% lift in organic clicks (Amsive). Branded queries with AIO present see CTR climb +18%. The CTR collapse is averaged across the whole SERP — but it’s not evenly distributed. The cited sources gain. Everyone else gets crushed harder than the headline number suggests. There is no middle.

Run the math on your own pipeline. If organic was sending 10,000 sessions per month at a 1.0% conversion rate, that’s 100 conversions. Cut sessions 58% and you’ve got 4,200. But if even a third of what remains is AI-referred at 4.4× conversion, you’re not at 42 conversions — you’re at 100+ again, with a fraction of the load on your funnel and customer support. The cost-per-conversion improves materially. This is why “losing clicks” can be a margin event, not a survival event.

What to do this week

First, instrument AI referrers properly. Most analytics setups still bucket Perplexity, ChatGPT, Copilot, and Gemini traffic into “Direct” or “Other” because of how the referrer header behaves. Build a regex segment for `perplexity.ai`, `chat.openai.com`, `chatgpt.com`, `copilot.microsoft.com`, and `gemini.google.com` in GA4 today. You cannot prove the 4.4× lift on your own funnel until you can see the source.

Second, pull a 60-day conversion-rate comparison: AI-referred sessions vs. classic organic. If you can’t pull it because volume is too low, that itself is the finding — you’re not yet visible in AI answers, which means the CTR collapse is hitting you with none of the offsetting upside. That’s the brief for the next quarter.

Third, kill any reporting deck that leads with sessions. Lead with revenue per channel and conversions per session, segmented by AI source. The CMO and CFO need a sentence like: “AIO took 58% of our SERP clicks, AI referrals replaced 70% of that revenue at higher margin, and we are investing in the citation work that closes the remaining 30%.” That sentence keeps the SEO budget alive. “Traffic is down” doesn’t.

Fourth, prioritize being the cited source. Front-load the answer, cite numbers, ship Q&A blocks with 40–60 word answers, and earn the Wikipedia/Wikidata + earned-byline footprint that gets you into the engines’ source pools. Cited brands take the +35% bump. Everyone else absorbs the -58%.

Need this done for you? Paris Roussos runs a flat-rate AI SEO service ($500–$1,500/mo per client, white-label for agencies) covering audits, schema and entity work, AI-visibility tracking, and content engineered to be cited by LLMs. Reach him at parisroussos@gmail.com.

Stop reporting traffic. Start reporting what the traffic is worth — and AI search starts looking like the best channel you’ve got.

A lot of sites quietly slammed the door on AI crawlers in late 2023 and 2024. It felt principled — “don’t train on our content without paying us” — and the one-line robots.txt edits were trivial. The bill came due the moment AI search took off. Pages that locked out GPTBot to protect themselves from training also vanished from the answer surfaces that now decide whether a brand gets cited. Most operators still haven’t separated those two decisions, and it’s costing them visibility in the only channel that’s growing.

The mechanic: training bots and answer bots share a parking lot

The AI labs run two distinct classes of crawler, and the names sit close enough that people conflate them. On one side: the training crawlers — GPTBot (OpenAI), ClaudeBot (Anthropic, when used for training), Google-Extended, CCBot, Bytespider, Amazonbot, Applebot-Extended, FacebookBot, anthropic-ai. These are the ones the publishing world reacted to, because their job is to vacuum content into model weights. On the other side: the answer-fetchers — OAI-SearchBot and ChatGPT-User (OpenAI), PerplexityBot and Perplexity-User (Perplexity), ClaudeBot when invoked for live browsing, and — quietly the most important of all — regular Googlebot, which is what feeds Google AI Overviews.

The training bots and the answer bots are different user agents. Blocking GPTBot does not opt you out of OpenAI’s search products, and blocking Google-Extended does not opt you out of AI Overviews. But — and this is where a lot of sites got cooked — most “block all AI scrapers” presets at the CDN layer (Cloudflare’s super-bot toggle, several WAF rule packs, half the WordPress security plugins) lump training and search agents into one bucket. You think you opted out of training. You actually opted out of being cited.

The stakes are now lopsided. AI Overviews queries that used to send organic traffic out are collapsing classic CTR by roughly 58–61% on informational intents, but the brands quoted inside the AI summary are seeing about a 35% lift in clicks on the cited surface. The pages that get cited are the pages the answer bots can fetch — full stop. If your robots.txt or your edge firewall says “no” to OAI-SearchBot, you are not in the running.

The two questions, decided separately

Stop treating “AI access” as one toggle. Decide twice.

Decision one — training. Are you OK with your content being used to train the next foundation model? Reasonable people land on either side. Block GPTBot, Google-Extended, anthropic-ai, CCBot, Bytespider if you want out of training. That’s a values/business-model call, not an SEO call.

Decision two — answer retrieval. Do you want to show up when someone asks ChatGPT, Perplexity, Claude, or Google AI Overviews about your category? If yes — and for almost every B2B and lead-gen brand the answer is yes — you must allow the answer-side user agents. Allow OAI-SearchBot and ChatGPT-User. Allow PerplexityBot and Perplexity-User. Leave regular Googlebot alone (it is your AIO pipeline). Allow ClaudeBot for live retrieval if you want Claude’s web mode to find you.

What to do this week

1. Pull your current robots.txt and read it with these two columns in mind: which user agents are blocked, and which class each one belongs to. If anything on the answer-fetch list is denied, that’s your bleeding wound.

2. Audit the edge. Cloudflare’s “block AI bots” managed rule, Sucuri’s AI scraper preset, similar WAF toggles — log into the dashboard and look at the actual UA list. The defaults change without notice and frequently bundle search bots into the training block.

3. Check your server logs for the past 30 days, filtered to OAI-SearchBot, ChatGPT-User, PerplexityBot, Perplexity-User, ClaudeBot. Zero requests is a red flag — either you blocked them, your CDN did, or your robots.txt is wrong. Don’t trust the documentation; trust the logs.

4. If you want explicit training opt-out language without losing citations, put the training bots in robots.txt with `Disallow: /`, leave the answer bots with `Allow: /`, and document the decision next to the file so the next person doesn’t “simplify” it.

Paris Roussos has been doing SEO since 1996 (co-founded a Forbes Best of the Web–winning site back in the day) and now runs a white-label AI SEO practice for agencies and brands — flat-rate, $500–$1,500/mo per client. If your top-of-funnel traffic is leaking into ChatGPT and Perplexity and you want it back, email parisroussos@gmail.com.

The right answer isn’t “block everything” or “allow everything” — it’s knowing exactly which bot you just told to leave.

If you’ve been treating the FAQ block at the bottom of your pages as a polite afterthought — a few collapsible questions tacked on for the long-tail keyword crowd — you’ve been sleeping on what is quickly becoming one of the most cited surfaces on the modern web. ChatGPT, Perplexity, Google AI Overviews, Gemini, and Claude are all disproportionately pulling answers out of question-and-answer blocks. Not because the FAQ block is magic. Because the format mirrors exactly how the LLM was asked the question in the first place.

The mechanic

When a user types “is X better than Y for Z” into ChatGPT, the model isn’t searching your page the way Googlebot would. It’s retrieving passages and ranking them by how cleanly they answer the question already in its context window. A paragraph that begins “Is X better than Y? Yes, when…” is a near-perfect match on the embedding side and a near-perfect match on the structural side. A 600-word essay that buries the same answer in paragraph six is two passes of summarization away from being useful.

That’s the cold mechanic underneath the FAQ effect. Modern AI search runs a passage-retrieval step before it runs any generation step, and passages that lead with the question form get retrieved at a much higher rate than passages that lead with a setup, a hook, or a brand mention. We already know from the citation data that 44.2% of LLM citations come from the first 30% of a page. Inside the first 30%, the format that gets quoted most consistently is the question-then-answer pair.

Schema makes it worse — or, rather, better, depending on which side of the deal you’re on. Pages that mark up Q&A blocks with `FAQPage` schema make it trivially easy for AI crawlers to parse the page into atomic answer units. Even when major AI bots aren’t formally honoring the markup (and most of them aren’t, as the llms.txt theater of the last year demonstrated), the HTML pattern itself — a heading framed as a question, followed by a short paragraph answer — is the most easily chunked structure on a webpage. Chunking is what retrieval cares about. Retrieval is what citation depends on.

What’s actually happening on cited pages

A practitioner pattern: the pages that get cited most by AI Overviews and Perplexity are not the longest, not the most beautifully designed, and not the ones with the most links. They’re the ones with the highest density of clean question-form headings followed by short, fact-dense answers. A page with twelve H2s shaped as questions — each followed by a 40-to-80-word answer that names entities, includes a stat, and ends without filler — outperforms a page with the same word count written as flowing prose. Every time.

This is also why competitor pages you didn’t think were competitive are suddenly showing up in AI answers. Someone published a “what is” / “how does” / “why do” page with a dozen Q&A pairs, and even with weaker domain authority, they’re getting cited because their page is shaped like the question.

What to do this week

Pick your top five pages by current Google traffic. Open each one and count the question-form headings. If the answer is fewer than three, you have work to do.

First, rewrite at least five H2s on each page as natural-language questions users would actually type into ChatGPT. “Pricing” becomes “How much does X cost in 2026?” “Features” becomes “What can X do that Y can’t?” Match the question form your buyer actually uses.

Second, front-load the answer in the first sentence after each question heading. One declarative sentence. No “Great question.” No setup. The model is looking for a passage that resolves the question in the first 30 words.

Third, add a dedicated FAQ block with five to ten question-answer pairs below the main content, marked up with `FAQPage` schema. Use questions you’ve actually seen in your support tickets, sales calls, and Reddit threads — not the ones a keyword tool suggested. Real questions get cited; SEO questions don’t.

Fourth, audit any existing FAQ block for filler. “It depends” answers, brand-promotional answers, and answers that don’t name a specific number, entity, or comparison don’t get retrieved. Rewrite them tight: name the thing, give the number, end the sentence.

Variant C — agency-targeted

Agencies: if your clients are starting to ask about AI SEO and you don’t have anyone in-house, Paris Roussos handles the work white-label — flat-rate, $500–$1,500/mo per end client, you keep the relationship. Email parisroussos@gmail.com for a sample audit.

The FAQ block stopped being a long-tail SEO trick the moment LLMs started doing the retrieval — now it’s the cheapest structural change you can make to your pages this quarter, and the one most likely to put your brand inside an AI answer next week.

Most marketing teams treat headings as visual furniture — bigger text to break up the page. That habit is now actively costing them AI citations. When a retrieval system decides whether a page is worth pulling into a ChatGPT, Perplexity, or AI Overview answer, the first thing it scans is the document outline. If the outline is sloppy, the page is filtered out before a single sentence is read.

A 2025 cross-engine analysis pinned the number down: 68.7% of pages cited inside AI answers follow a strict H1 → H2 → H3 hierarchy. The non-cited control group looked nothing like that. It was a mess of skipped levels, double H1s, decorative H4s, and headings nested for visual weight rather than meaning. Same words. Same content. Different fate inside the LLM.

If you’ve been doing SEO long enough to remember when H1s were “the rule,” this is going to feel like 2010 came back wearing a new badge. It kind of did.

The mechanic: LLMs parse hierarchy the same way they parse a knowledge graph

Retrieval pipelines for ChatGPT, Perplexity, Gemini, and Google’s AI Overviews don’t read your page top-to-bottom like a person. They chunk it. The chunker uses headings as the segment boundaries. Your H2 isn’t a design choice — it’s a function call that says, “the next ~200 words are an answerable unit about X.”

When the hierarchy is clean, the model gets a clear map: H1 names the entity, each H2 names a subtopic, each H3 narrows the subtopic into a concrete claim. The retriever scores those chunks against the user’s query and picks the tightest match.

When the hierarchy is broken — H1, H3, H2, H4, H2 — the chunker still segments, but the segments don’t align with concepts. The model can’t tell whether your H4 is a sibling of the previous H3 or a stray label. Confidence drops. Citation probability drops with it. This is the same failure mode you used to see in featured-snippet eligibility, just with bigger consequences and no second chance.

Pair this with the front-load rule (44.2% of LLM citations come from the first 30% of the page) and the picture sharpens fast: the model is trying to find a confident, top-of-page, well-bounded answer block. Headings are the bounding boxes.

Why most pages quietly fail the test

I audit a lot of agency client sites, and the same four patterns kill heading hierarchy in production:

1. Double H1s. A logo or site name is wrapped in an H1, and the actual article title also gets an H1. Two H1s, no clear root. The retriever now has to guess which one names the entity.

2. Visual-weight nesting. Designers reach for H3 because it “looks right” between paragraphs, even when the previous heading was an H1. The outline jumps from level 1 to level 3 with nothing in between. The chunker sees a missing parent.

3. Decorative H4s. Sidebars, callout boxes, and “related reading” widgets ship with H4s that have nothing to do with the article’s structure. They contaminate the outline.

4. Page-builder defaults. Elementor, Divi, Gutenberg blocks, half the React component libraries — they all have at least one preset that emits a wrong-level heading. Nobody catches it because the page looks fine.

None of this shows up in a Lighthouse score. It shows up in your AI citation rate, which most teams aren’t tracking yet, so the bleeding is invisible.

What to do this week

• Pull your top 20 traffic pages and run them through any free outline auditor (W3C validator, axe DevTools, or just `document.querySelectorAll(‘h1,h2,h3,h4,h5,h6’)` in the console). Look for skipped levels and multiple H1s.
• Lock the rule for new content: one H1 per page, H2s for major sections, H3s only as children of an H2, H4 used sparingly or never. If you can’t justify an H4, delete it.
• Audit your global components — header, sidebar, post meta, related posts, footer — for stray heading tags. Demote them to `

  ` or `

  ` if you need the styling.
• Have your writers draft in outline form first (H1, H2s, H3s only), then fill in paragraphs. Pages drafted that way pass the hierarchy test by construction.

This is unglamorous work. It’s also one of the highest-leverage on-page changes you can make right now, because most of your competitors haven’t connected the heading-hierarchy dot to AI citation outcomes yet. They will. The window is open while it’s still considered an accessibility nice-to-have instead of an AI visibility lever.

Need this done for you? Paris Roussos runs a flat-rate AI SEO service ($500–$1,500/mo per client, white-label for agencies) covering audits, schema and entity work, AI-visibility tracking, and content engineered to be cited by LLMs. Reach him at parisroussos@gmail.com.

Clean outlines won citations in 2010. They’re winning them again — just with a much bigger payout.