How to structure a blog post so AI search cites it

Most Shopify blog posts are written to rank, not to be quoted. That distinction now costs money. When ChatGPT, Perplexity, or Google AI Overviews answer "what are the warmest socks for winter hiking", they pull a handful of passages from the web, paraphrase them, and cite the pages those passages came from. Your post is either one of those passages or it does not exist in that answer.

This guide walks through the anatomy of a citable post, section by section: the opening answer, question-form headings, quotable sentences, comparison tables, FAQ blocks, JSON-LD, length, freshness, and internal links. Every rule carries a number from published research or a first-hand observation from our own weekly citation tracking. To keep it concrete, one illustrative example runs through the whole guide: a made-up merino sock store called Ridgeline Socks. It is not a client, and its product numbers are invented for the example. The structure is the point, and it transfers to cast-iron pans, skincare, or dog food without changes.

What does a citable blog post look like?

A citable blog post answers the target question in its first paragraph, breaks the rest into question-form sections that each open with one quotable sentence, and backs those sentences with statistics, quotations, named sources, a comparison table, a FAQ block, and JSON-LD that mirrors the visible text.

That is a long sentence for a short idea: the page is built from passages, and each passage is built to be lifted out. Generative engines do not read your post the way a customer does. They retrieve candidate pages, split them into chunks, score the chunks against the prompt and its sub-queries, and hand the winners to a language model that writes the answer and cites where each claim came from. The mechanism is covered in detail in how ChatGPT and Perplexity choose which pages to cite. What matters here is the consequence: structure decides which chunk of your page is even eligible.

At a glance, the checklist for one post:

• A title that contains the buyer's question, under 70 characters.
• A direct answer in the first 80 words, with at least one number.
• 8 to 12 H2 sections, at least 6 phrased as questions.
• One quotable first sentence per section, then evidence.
• One comparison table or "at a glance" list per major decision.
• A FAQ block of 5 to 8 self-contained answers.
• BlogPosting plus FAQPage JSON-LD that repeats the visible text.
• A visible updated date, and links to the relevant products and collection.

The rest of this guide takes each block in turn.

Why do the first 80 words decide whether you get cited?

The first 80 words decide the citation because generative engines extract passages, not pages, and the passage they test first is the one directly under your title.

The GEO paper from Princeton and IIT Delhi (Aggarwal et al., published at KDD 2024) describes the standard design: the engine fetches the top sources for a query, then a language model writes the answer from those sources. In their test setup, only the top five sources were fetched per query. Google says its AI Overviews and AI Mode use a "query fan-out" technique, issuing multiple related searches across subtopics, per its AI features documentation. In both cases the model is working from a small window of text per source. If your window opens with a warm-up paragraph, the model has nothing to quote.

Here is the difference for Ridgeline Socks on the question "what are the warmest socks for winter hiking?"

Weak opening: "Winter is coming, and with it the age-old question every hiker faces: how do I keep my feet warm on the trail? At Ridgeline we have spent years obsessing over this, and in this post we share everything we have learned."

Citable opening: "The warmest socks for winter hiking are heavy-weight merino blends: 60 to 80% merino, 20 to 40% nylon, a cushioned sole, and a crew height that clears the boot cuff. Below freezing, choose a fabric weight of 300 g/m2 or more. This guide compares three weights by temperature range and explains why nylon content matters more for durability than merino percentage." (Numbers invented for the example.)

The second version is 62 words. It names the answer, gives two thresholds, and previews the structure. An extractor can quote the first sentence on its own. A buyer reading it knows within five seconds whether to keep going. Both audiences are served by the same paragraph, which is the whole idea behind this structure.

How should you phrase H2 headings so they match AI prompts?

Phrase each H2 as the exact question a buyer would type into ChatGPT, because engines match your headings and titles against the prompt and its fan-out sub-queries before they read the body.

Ahrefs analyzed 1.4 million ChatGPT prompts and compared pages ChatGPT retrieved and cited with pages it retrieved and ignored. Cited pages had a title-to-prompt cosine similarity of 0.602 against 0.484 for non-cited pages. Pages with natural-language URLs were cited 89.78% of the time versus 81.11% for opaque slugs. Headings are the same signal one level down: each H2 is a mini title for its passage, and a question-form heading matches a question-form prompt more closely than a label does.

For the Ridgeline guide, the H2 set looks like this:

• How warm are merino socks compared to synthetic socks?
• What merino percentage is best for hiking socks?
• Which sock weight should you wear at which temperature?
• How long do merino hiking socks last?
• Do merino socks stop blisters?
• Are merino socks worth $28 a pair?
• How should you wash merino socks?
• What to do next

Seven of eight are questions. None repeats the title. Each one is a query someone actually types, including the price objection. Keep headings under 12 words, avoid brand names inside them (the brand belongs in the answer, not the question), and do not stack synonyms to catch variants. One clean question per section is what the fan-out matches.

What makes a sentence quotable enough to be extracted?

A quotable sentence states one specific claim in under 25 words, contains a number or a named source, and still makes sense with the rest of the page removed.

This is the part of the structure with the strongest evidence. The GEO paper tested nine ways of rewriting a source and measured how much of the generated answer each version earned. Its authors report that the three top methods, Cite Sources, Quotation Addition, and Statistics Addition, "achieved a relative improvement of 30-40% on the Position-Adjusted Word Count metric and 15-30% on the Subjective Impression metric." Keyword Stuffing, the classic SEO move, "performs 10% worse than the baseline." The absolute scores from the paper's main results table:

Source: Aggarwal et al., GEO-bench (10,000 queries), Table 1. Higher is more visible in the generated answer.

Two more findings from the same paper matter for a small store. First, when the authors ran their methods against Perplexity.ai as a live engine, Cite Sources and Statistics Addition still showed improvements of up to 9% and 37% on the two metrics. Second, the lift was largest for pages that were not already on top: Cite Sources raised the visibility of the fifth-ranked source by 115.1% while the top-ranked source lost 30.3%. A Shopify blog is rarely the top result for a category question. Specific, attributed sentences are how the fifth result gets quoted anyway.

For Ridgeline, "quotable" means writing claims the store can stand behind: fabric weights in g/m2, temperature ranges it recommends, wash-cycle results from its own tests, and quotations from named third parties where they exist (a textile lab, a standards body, a published review). Where a number is not available, the fallback is a named mechanism ("nylon carries the abrasion load at the heel and toe"), not an adjective.

When do comparison tables and at-a-glance lists earn a citation?

Comparison tables and short lists earn citations when the question has a discrete answer set, such as sizes, weights, prices, or feature differences, because an engine can lift one row without rewriting it.

We should be honest about the evidence here. No published study we have found isolates tables as a citation factor across ChatGPT, Perplexity, and AI Overviews. What exists is adjacent: the GEO paper's Statistics Addition result, and our own tracking. Citevolt re-queries priority keywords for each client weekly across the three engines. On a baseline run on 2026-09-14 across six buyer queries, every cited page was a long explainer or comparison article, and none was a product page or a short listicle. Comparison content gets cited; the table is the most extractable form of it.

The Ridgeline post has one obvious table, the weight-to-temperature guide:

(Illustrative values for a made-up store. Yours must come from your product data.)

Three rules keep tables citable. Put the decision variable in the first column. Use units in the header, not in every cell. Keep the table to 2 to 5 columns and 3 to 8 rows so it survives a narrow viewport and a passage window. An "at a glance" bullet list right after the opening answer does the same job for non-tabular decisions: three to six bullets, each a complete sentence, each carrying one fact.

Do FAQ blocks still matter now that Google dropped FAQ rich results?

FAQ blocks still matter for AI citation because each self-contained answer is a ready-made passage, even though Google stopped showing FAQ rich results on May 7, 2026.

The timeline is worth knowing so you do not build for a feature that is gone. On August 8, 2023, Google reduced FAQ rich results to a narrow set of sites: "Going forward, FAQ (from FAQPage structured data) rich results will only be shown for well-known, authoritative government and health websites." In May 2026 Google ended the feature entirely, dropping the search appearance and Rich Results Test support in June and the Search Console API support in August. Google's own guidance was that publishers may leave the markup in place or remove it.

So why keep the block? Because the visible FAQ was never really about the dropdown. It is the section of your post where each question gets a 40 to 90 word answer that makes sense with nothing around it, which is exactly what a retriever wants. FAQ answers cover the adjacent questions that do not deserve a full H2: sizing, care, returns, shipping to a specific country, compatibility with a specific boot. For Ridgeline, that is "do merino socks shrink in the dryer?" and "what size should I order if I am between sizes?" Each answer states the rule, gives the number, and stops.

Two constraints. Do not paste body paragraphs into the FAQ; write new answers to new questions. And do not let the FAQ turn into marketing ("why is Ridgeline the best?"). Engines cite answers to questions people ask, and nobody asks that one.

Which JSON-LD should a Shopify blog post carry?

A Shopify blog post should carry one BlogPosting block with headline, datePublished, dateModified, author, and publisher, plus one FAQPage block that repeats the visible FAQ word for word.

Start with what Google says. Its AI features page states: "You don't need to create new machine readable files, AI text files, or markup to appear in these features." That is accurate and it is not a reason to skip structured data. In the May 2025 guidance on AI search, John Mueller lists "make sure structured data matches the visible content" as one of the core recommendations, with the rule that all content in your markup must also be visible on the page. Structured data is how you tell every crawler, including OAI-SearchBot and PerplexityBot, what the page is, who wrote it, and when it changed, in a form that does not depend on parsing your theme's HTML.

Google's Article structured data documentation lists the recommended properties: headline, image, datePublished, dateModified, and author, with dates in ISO 8601. On schema.org, BlogPosting sits under Article, and FAQPage is a WebPage holding Question and Answer pairs. A minimal, honest block for the Ridgeline post looks like this:

{
  "@context": "https://schema.org",
  "@graph": [
    {
      "@type": "BlogPosting",
      "@id": "https://example.com/blogs/guides/warmest-socks-winter-hiking#article",
      "mainEntityOfPage": "https://example.com/blogs/guides/warmest-socks-winter-hiking",
      "headline": "What are the warmest socks for winter hiking?",
      "description": "Heavy-weight merino blends with 20 to 40% nylon, a cushioned sole and crew height are the warmest choice. This guide compares three weights by temperature range.",
      "image": "https://example.com/cdn/warmest-socks-hero.jpg",
      "datePublished": "2026-09-14T08:00:00+01:00",
      "dateModified": "2026-09-14T08:00:00+01:00",
      "author": { "@type": "Person", "name": "Founder name" },
      "publisher": {
        "@type": "Organization",
        "name": "Ridgeline Socks",
        "logo": { "@type": "ImageObject", "url": "https://example.com/logo.png" }
      },
      "about": {
        "@type": "Product",
        "name": "Ridgeline Heavyweight Merino Crew",
        "url": "https://example.com/products/heavyweight-merino-crew"
      }
    },
    {
      "@type": "FAQPage",
      "mainEntity": [
        {
          "@type": "Question",
          "name": "Do merino socks shrink in the dryer?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "Yes, on high heat. Wash cold, tumble dry low or air dry, and a merino-nylon blend holds its size. Pure merino is more sensitive."
          }
        }
      ]
    }
  ]
}

Where it goes on Shopify: in the article template of your theme, or inline in the article body HTML. Many themes already emit a basic Article block, so view the page source and search for application/ld+json before adding a second one; two Article blocks with different dates are worse than one. Validate with Google's Rich Results Test or the schema.org validator before you publish. Citevolt injects both blocks into every post it publishes and checks them, which is one of the things the $297 to $697 per month plans cover; the point of this section is that you can do it by hand for a single post in ten minutes.

How long should the post be?

Length does not predict citation on its own: across 174,048 pages cited in AI Overviews, Ahrefs measured a 0.04 Spearman correlation between word count and citation, and 53.4% of cited pages were under 1,000 words.

The Ahrefs study (December 2025) is the largest public dataset on this question. The average cited page ran 1,282 words. Articles specifically had a median of 1,166 words. 30.6% of cited pages fell between 1,000 and 2,000 words, and 16% ran over 2,000. Pages in citation positions 1 to 3 averaged around 1,270 to 1,291 words, while positions 4 to 10 averaged 1,690, so longer pages were slightly more likely to be cited lower in the list, not higher. Two caveats: the study covers Google AI Overviews only, not ChatGPT or Perplexity, and 76% of AI Overview citations come from pages already in Google's top 10 per Ahrefs' statistics roundup, so ranking factors are entangled with length in that data.

Our first-hand observation cuts the other way for buyer queries. The cited pages in our September 2026 baseline were long explainer and comparison articles, typically well over 1,500 words, because a comparison across three weights, five competitors, and eight questions takes that much text. The reconciliation is that length is a byproduct of coverage. A post that answers one narrow question in 900 words with a table and a FAQ is citable. A 3,000-word post that circles the same point is not.

Practical ranges: 1,200 to 2,000 words for a single product question, 2,000 to 3,500 for a comparison or buying guide. Citevolt sets each post's target at 1.1 times the average word count of the top three organic results for the keyword, within the plan's range (1,200 to 1,500 words on Starter, 1,500 to 3,500 on Growth, 2,000 to 3,500 on Scale). Write until every sub-question is answered with a specific claim, then stop.

How much do freshness and dateModified matter?

Freshness matters measurably: across 17 million citations, Ahrefs found AI assistants cite pages averaging 1,064 days old versus 1,432 days for organic results, 25.7% fresher.

That Ahrefs freshness study (July 2025) also broke it down by engine: ChatGPT citations averaged 958 days old, Perplexity 1,166, Google AI Overviews 1,432. Seer Interactive's recency study of 5,000-plus cited URLs (June 2025) found about 50% of Perplexity's citations pointed to content published in 2025, versus roughly 44% for AI Overviews and 31% for ChatGPT. Ahrefs' separate ChatGPT analysis put the median age of a cited page at around 500 days, and its statistics roundup reports that 76.4% of ChatGPT's most-cited pages had been updated in the previous 30 days.

Read together: the engines still prefer pages that have been around for a while, and they prefer the ones among those that have been touched recently. A post published in 2024 and refreshed this quarter is in the best position. A post published in 2024 and never touched is drifting out.

What that means for structure:

• Show an "Updated on" date near the top of the post, in the visible text, not only in the metadata.
• Keep datePublished fixed and move dateModified in the JSON-LD only when the visible content changes, per Google's rule that markup matches the page.
• Refresh the numbers, the product links, and the FAQ, not the headline. Rewriting the title resets the signal that the page is established.
• Put refresh on a schedule. Two to five content refreshes per month is the cadence Citevolt builds into its Growth and Scale plans, and a quarterly pass over your top 20 posts is a reasonable manual equivalent.

Where should internal links to products and collections go?

Put the first product link inside the section that answers the buying question, and the collection link in the closing "what to do next" section, so a reader who arrives from an AI answer can act within one click.

Internal links do two jobs on a Shopify blog. For the buyer, they turn a citation into a visit to a product page. For the engine, they tell it which product entity the post is about, which is why the about property in the JSON-LD above points at a product URL. What they do not do is substitute for evidence. Ahrefs' 75,000-brand correlation study (December 2025) found brand web mentions correlated with ChatGPT visibility at 0.664, while the number of pages on a site correlated at roughly 0.194. Publishing more pages, or linking them to each other more densely, is not what moves visibility. Being cited for specific claims is.

For the Ridgeline guide, the link plan is: the heavyweight crew sock linked from the weight table row that recommends it, the mid-weight sock linked from the "which weight at which temperature" section, the winter collection linked once in "what to do next", and one link to the store's care guide from the wash section. Four links, each with an anchor that describes the destination ("Ridgeline Heavyweight Merino Crew, 320 g/m2"), none with "click here" or "shop now." The guide to getting a Shopify store cited by ChatGPT covers the crawl and access side, including which bots need to be allowed for these links to be followed.

What should you avoid?

Avoid anything that delays or dilutes the answer: warm-up intros, hedged claims, unsourced numbers, keyword repetition, and paragraphs longer than five sentences.

Each of these has a specific cost.

• Fluff intros. The passage under your title is the one that gets scored first. A 150-word scene-setter means the first scored passage has no answer in it.
• Hedging. "Some experts say merino may be warmer for many people" contains nothing to extract. State the claim, give the condition under which it holds, cite the source.
• Unsourced claims. Cite Sources was one of the three top-performing methods in the GEO benchmark. An unattributed number reads as an opinion to a model that is deciding what to attribute.
• Keyword stuffing. Measured at 10% below the unmodified baseline in the same benchmark. Repeating "best merino hiking socks" in every heading makes the page less citable, not more.
• Walls of text. A retriever chunks by structure. A 400-word paragraph is one chunk with six ideas in it, and the score for the one idea that matches the prompt is dragged down by the five that do not.
• Marketing in the FAQ. Nobody prompts "why is this brand the best?" so no engine cites the answer.

The before-and-after view for the same Ridgeline topic:

The left column is what most Shopify blogs look like today, which is the opportunity.

What to do next

Take one post, the one closest to a question your buyers actually type, and rebuild it in this order.

1. Pick the target question and put it in the title, under 70 characters. Check it against what your customers ask in support tickets and reviews.
2. Write the 80-word direct answer first, before any other section. Include at least one number or threshold.
3. Draft 8 to 12 H2 headings, at least 6 as questions. Reject any heading a buyer would not type.
4. Write the first sentence of each section to pass the 25-word test, then add the evidence: a statistic, a quotation, or a named source. Invent nothing.
5. Add one comparison table for the main decision and a FAQ of 5 to 8 adjacent questions with self-contained answers.
6. Add BlogPosting and FAQPage JSON-LD that repeats the visible text, check for a duplicate Article block from your theme, and validate it.
7. Link the relevant products from the sections that recommend them and the collection from this step. Use descriptive anchors.
8. Publish with a visible updated date. Re-run the target question in ChatGPT, Perplexity, and Google four weeks later and note which pages are cited.

If you want to see where your store stands before rebuilding anything, the free Citevolt audit runs your category's buyer queries through all three engines and shows which pages are being cited instead of yours. If you would rather not do the rebuild by hand, Citevolt writes and publishes posts in this structure to your Shopify blog weekly, from $297/mo for 4 posts; what that buys compared with agencies and tools is covered in AI SEO agency pricing: what AEO services cost. Either way, the structure above is the deliverable. Everything else is cadence.