How to Get Cited by ChatGPT and AI Overviews

Getting cited by ChatGPT and AI Overviews is decided at the paragraph level, not the domain level. A study of 1,000 AI Overviews collected in April 2026 found that the top 1% of domains — roughly twelve sites — capture 47% of all citations, while the remaining 90% of domains share 22%; the same study found that pages citing at least one named source inline in the body were cited 2.1× more often than pages that didn't. That second number is the one a 20-post niche blog can actually move. You will not out-cite Wikipedia, but the half of the citation market that isn't already spoken for is won by pages that are easy to lift a self-contained, evidence-dense passage out of — and that is a writing decision, not a link-building budget.

Below is what actually holds up: the one controlled experiment, the crawler logs, the citation datasets that disagree with each other, and the technical floor that quietly disqualifies small sites before any writing advice matters.

The door is only half shut

Start with the ceiling, because pretending it isn't there makes everything after it less credible.

DigitalApplied analysed 1,000 AI Overviews — 100 queries across 10 intent classes, roughly 30 verticals, US-English desktop, collected 8–22 April 2026 — crawling 4,243 unique cited URLs against about 50,000 non-cited controls. The distribution is brutally top-heavy: the top 1% of domains capture 47% of citations, the next 9% capture 31%, and the long-tail 90% split the remaining 22%. Their own summary is the sentence worth pinning above your desk: "Twelve domains capture half of all AIO citations. The other half is the only addressable market for everyone else — and it is decided by page-level signals, not domain age."

Now the flip side, from the only controlled experiment anyone has run on this.

The GEO paper out of Princeton and IIT-Delhi (KDD 2024) built GEO-bench, a benchmark of 10,000 queries run through a two-stage generative engine: retrieve the top five sources, then synthesise a cited answer. The researchers tested nine content modifications and measured how much of the final answer each source contributed. The headline result is well-travelled — the paper reports visibility gains "up to 40%" — but the more interesting number is buried in the rank breakdown.

Applying the "Cite Sources" method (adding relevant citations to claims in the text) to the source ranked 5th in the retrieved set lifted its visibility by +115.1%. Applying the exact same method to the top-ranked source lowered its visibility by 30.3%. The paper frames this explicitly as GEO's democratising potential for smaller creators competing against established domains.

Carry two caveats: GEO-bench ran on GPT-3.5-turbo-era engines and nobody has publicly re-run it on 2026 frontier models, and "up to 40%" is a maximum on one metric, not an average. Keyword stuffing was among the worst-performing methods tested, so this is no licence for tricks. Treat it as directional: attribution-dense prose helps most when you are not already the top result. Which describes every small blog.

What actually changed: you are competing for a passage, not a page

The mechanism is not a secret. Google described it in its own words when it shipped AI Mode: "Under the hood, AI Mode uses our query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf."

That single sentence rearranges the job. Your page is not being ranked against competitors for a keyword. It is being retrieved as one of several candidates for a sub-question you probably never targeted, then read for a passage that answers that sub-question cleanly.

An April 2026 measurement study (arXiv 2604.25707) makes the useful distinction here. Across 602 controlled prompts, 21,143 valid search-layer citations and 23,745 citation-level feature records spanning ChatGPT, Google AI Overview/Gemini and Perplexity, the authors separate citation selection from citation absorption — whether your page appears in the source list, versus whether your text actually shapes the answer's wording, reasoning and facts. Different games. Pages with high absorption shared three traits: extended length and structural organisation, strong semantic alignment with query intent, and rich extractable evidence — which the paper names as "definitions, numerical facts, comparisons, and procedural steps."

Breadth and depth also diverge by engine: Perplexity and Google reference more sources overall, while ChatGPT cites fewer but shows substantially higher average citation influence per fetched page. If ChatGPT is your target, being one of the few sources selected matters more than being one of many.

The ranking connection is loosening too, though the data needs care. Ahrefs' July 2025 study of 1.9 million citations across 1 million AI Overviews found 76.10% of cited pages ranked in the top 10 and 14.40% ranked below position 100 with no SERP visibility at all. Their March 2026 re-run across 863,000 keywords and 4 million AI Overview URLs put the top-10 share at just 38%, with 31.2% from positions 11–100 and 31.0% from beyond position 100. Ahrefs attributes part of that shift to improved parsing methodology, so the two studies are not directly comparable — but even the conservative reading says a meaningful share of citations come from pages that would never win the blue link. BrightEdge, measuring across nine industries from May 2024 to September 2025, found only 16.7% of AI Overview citations come from top-10 results and called positions 21–100 "the sweet spot."

Ahrefs' Si Quan Ong put the honest version best: "It's true that if you rank #1 in the SERPs, you're more likely to be cited in an AI Overview than if you were ranked lower. But that chance is a coin flip at best."

The four things AI answers reach for

The absorption study's list — definitions, numerical facts, comparisons, procedural steps — is the most useful finding in this space, because all four are paragraph-shaped. You don't need a domain, a budget or a decade. You need to put one of those four things in a chunk of text that survives being torn out of the page.

Cross-reference it with GEO-bench's per-method results on Position-Adjusted Word Count — how much of the generated answer a source contributes, weighted by position — and the four best-performing methods were:

Method	Lift on Position-Adjusted Word Count	Lift on Subjective Impression
Quotation Addition	+41%	+28%
Statistics Addition	+33%	+23%
Fluency Optimization	+29%	+13%
Cite Sources	+28%	+14%

Three of those four are forms of attribution. That lines up with DigitalApplied's page-level finding that pages citing a named source inline were cited 2.1× more often — measured on 4,243 cited URLs against ~50,000 controls, so correlational, not causal, but pointing the same direction as the experiment.

What that looks like in an actual paragraph

Here is the shape, using an illustrative rewrite rather than a real page.

Before:

There's been a lot of discussion lately about whether domain authority still matters for AI search. Many experts believe it's becoming less important, though opinions vary. We'll explore what this means for smaller publishers below.

That paragraph is unliftable. It contains no definition, no number, no named source, and no claim. Pulled out of the page it says nothing.

After:

Domain authority barely predicts AI citation. Surfer analysed roughly 20,000 prompts and about 5 million unique citation source URLs across Google AI Mode, AI Overviews, OpenAI and Perplexity, and found Spearman correlations of −0.065 for PageRank and −0.053 for Domain Score against AI Overview citation frequency — what they described as "close enough to zero to be noise." The relationship held near zero even after removing the top 5% of domains.

Same length. The second version has a committed claim in the first sentence, a named source, a sample size, two numbers with their metric named, and a direct quote. Torn out of the page, it still works. That is the entire test.

The pattern, minus the fake numbers

The dominant vendor prescription for this is the "answer capsule": a 40–60 word self-contained answer placed directly under an H2, structured as direct answer → qualifier → scope or exception, in plain paragraph prose, third person, with no hyperlinks inside the capsule itself. The named failure modes are preamble openers, link-laden capsules, hedging language, callout formatting, and capsules containing no numbers.

The pattern is sensible and matches what the absorption study measured independently. One integrity note, since this audience will check: the statistics that page and its peers attach to the pattern (things like "72.4% of cited pages use answer capsules") chain-cite other GEO vendor blogs rather than any traceable dataset. Use the pattern. Don't repeat the numbers.

One case study is worth reading for shape rather than proof: Growtika's, describing a bootstrapped client competing against companies that had raised $100M+. The on-page work was unglamorous — 40–50 word answer blocks structured as direct answer → supporting details → differentiator, data tables with specific pricing and features, and comparison pages with exact numbers — and the reported outcome was 172 citations across ChatGPT and Claude, third behind only Wikipedia and arXiv. The number to internalise is the timeline: zero citations for two months, first long-tail citations at months 3–4, consistent cross-platform citations at months 9–10. It's an agency's own unaudited case study with an anonymous client — but the year-long arc matches everything else here.

Cite correctly, or don't cite

If the highest-leverage thing you can do to a paragraph is attach a named source to a number, the quality of that attribution becomes load-bearing. A dead or invented link is worse than no link — it costs a reader's trust the first time they click, and it's the most common failure mode of AI-drafted content.

That's the part of the workflow Magic Share's link verification and fact-checking is built around: a draft isn't saved until every cited URL resolves, and a verification report ships attached so review takes a coffee rather than an afternoon. "Add more attributed statistics" is only good advice if the attributions hold.

Authority matters less than you'd think — and mentions matter more

This section comes with a warning label, because the two best datasets flatly disagree.

Surfer's study — roughly 20,000 prompts and about 5 million unique citation source URLs across AI Mode, AI Overviews, OpenAI and Perplexity, over roughly three months — found authority metrics correlate negatively but negligibly with citation frequency: AI Overviews at −0.065 (PageRank) and −0.053 (Domain Score) on full-dataset Spearman. Strip out the top 5% of domains and everything collapses to within ±0.05 of zero. DigitalApplied, on its much smaller 4,243-URL sample, found the opposite: a +0.61 Pearson correlation between its domain authority proxy and citation rate.

Neither is peer-reviewed. Surfer's dataset is roughly a thousand times larger; DigitalApplied is more transparent about its controls. The honest synthesis is not "authority doesn't matter" — it is that authority is not the gate in AI citation that it is in classical rankings, and it explains far less of the variance than practitioners assume.

Ahrefs' 75,000-brand study from May 2025 splits the difference in a way that's more actionable than either extreme. Running Spearman correlations against AI Overview brand visibility, it ranked branded web mentions highest at 0.664, branded anchors at 0.527, branded search volume at 0.392, Domain Rating at 0.326, referring domains at 0.295, and raw backlinks at 0.218. Author Louise Linehan is explicit that "while the data shows statistical relationships, I should emphasize that correlation ≠ causation."

Read that table as a small publisher and it says something specific: being talked about in text correlates about three times more strongly than being linked to. A mention in a forum thread or a newsletter is something a niche blog can earn with one genuinely useful piece of research. A backlink profile is not.

Gianluca Fiorelli, quoted in Surfer's study, described the target: "The winning position — for both traditional rankings and AI citations — is content anchored in established consensus that simultaneously contributes genuine new information." A decent one-sentence brief for every post you write from here on.

Dates, updates, and the freshness question

Seer Interactive tracked 7,683 dated pages carrying 47,097 citations across ChatGPT, Gemini and Perplexity for four brands between March and June 2026. Seventy-five percent of cited pages had been updated within the past year and 88% within two years; content older than three years was, in their words, "basically non-existent." By engine: Gemini 78% within one year, ChatGPT 73%, Perplexity 65%.

The split that matters for a small library is this one: by last-update date, 72% of cited pages look fresh — but by original publish date, only 42% do. Twenty-seven percent of the "fresh" cited content was originally published more than two years ago. Sonny Vasquez of Seer summarised it as "The freshness LLMs reward is being manufactured by updates, not by new publishing," and added the corollary: "Newest gets you the spike, established plus maintained gets you the staying power."

New content gets the burst; maintained content gets the annuity.

And here is the conflict, stated plainly rather than resolved conveniently: DigitalApplied's sample found no correlation between page recency and citation, with a median cited page age of 14 months. Both findings can be true at once — 14 months sits comfortably inside "updated within two years." The defensible advice that survives both datasets is narrow and cheap to implement: keep a visible last-updated date on the page, and make sure it reflects an actual update. Not a publishing treadmill.

For a 20-post blog this is good news, because it reframes the job. You are not competing on volume against a content team. You are maintaining a small library so each page stays accurate, stays dated, and stays worth quoting. That is what "a blog compounds when it's tended year-round, not in bursts" means in citation terms — and it's the reasoning behind Magic Share's auto-draft scheduling and topic ledger: the agent wakes at 06:00 UTC, drafts what's due on your cadence, checks the ledger so ideas never repeat, and waits for you to approve.

The technical floor that quietly makes a blog uncitable

None of the paragraph craft matters if the crawler can't read the page. This is the part most GEO advice skips, and the part a technical founder can fix in an afternoon.

AI crawlers do not run JavaScript. Vercel and MERJ analysed a month of crawler traffic across their network — Googlebot 4.5B fetches, GPTBot 569M, ClaudeBot 370M, AppleBot 314M, PerplexityBot 24.4M — and found none of GPTBot, ClaudeBot or PerplexityBot execute JavaScript. GPTBot fetched JS files in 11.50% of requests and ClaudeBot in 23.84%, and executed none of them. AppleBot and Google's Gemini do render. If your blog is client-side rendered, most AI crawlers see an empty shell.

They also pick URLs badly. In the same study, ChatGPT's crawler spent 34.82% of its fetches on 404s and Claude's 34.16%, versus 8.22% for Googlebot; GPTBot spent a further 14.36% following redirects. A clean sitemap, stable URLs and no dead internal links matter more for AI crawlers than for Google, not less.

There's a mundane advantage here for anyone publishing markdown to a static site: server-rendered HTML at a stable URL is exactly the shape these crawlers can read, by construction. That's a property of the publishing target rather than of any tool — but a workflow where posts land as merged pull requests starts on the right side of this problem without anyone thinking about it.

Check your OpenAI bot access separately from your training-bot policy. OpenAI's docs distinguish three crawlers: OAI-SearchBot surfaces content in ChatGPT search, GPTBot collects training data, and ChatGPT-User handles user-initiated fetches. The consequential line is explicit: "Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers." Blocking GPTBot to opt out of training is a defensible choice; blocking OAI-SearchBot removes you from ChatGPT search entirely — and plenty of sites have done the second while intending only the first.

Check your snippet directives too. Google's documentation states that to be eligible for generative AI features, a page must be indexed and eligible to be shown with a snippet, with noindex, nosnippet, data-nosnippet, max-snippet and Google-Extended as the available controls. A privacy-minded nosnippet or an aggressive max-snippet opts you out of AI Overviews, quietly.

Cloudflare's defaults change on 15 September 2026. In an announcement dated 1 July 2026, Cloudflare defined three crawler categories — Search ("collects or indexes your content, so it can answer questions about it later"), Agent, and Training — and said that from 15 September 2026, Training and Agent bots will be blocked by default on ad-monetised pages, with Search bots still allowed. It applies to newly onboarded domains, new customers and existing Free-tier customers, with an opt-out window. If you're on the free tier with ads, check which side of that line your pages sit on before September.

Otterly's January–February 2026 analysis of over a million citations reports that 73% of websites have technical obstacles blocking AI crawler access — robots.txt rules, CDN security settings, or JavaScript rendering dependencies. Treat that percentage as indicative, but the failure modes it names are the same ones Vercel, OpenAI and Cloudflare document independently.

The one file you can skip

Ahrefs checked 137,210 domains in May 2026 using server-log data. Twenty-eight percent publish an llms.txt file. Ninety-seven percent of those files received zero requests that month. Only 19.5% of the fetches that did arrive came from named AI tools, and zero AI bot requests targeted llms.txt files that don't exist — meaning AI bots aren't probing for it. As Ahrefs put it: "it is not a robots.txt-style directive: it controls nothing and blocks nothing." Gary Illyes has confirmed Google doesn't support it and has no plans to, and John Mueller compared it to the dead meta keywords tag: "It's basically you're telling these systems, like, I have the best website ever."

Which brings up the objection this audience will raise next.

"But Google says none of this is necessary"

It does. Google's official guide to optimizing for generative AI features states, verbatim: "You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities)." Also: "There's no requirement to break your content into tiny pieces for AI to better understand it." And most bluntly: "You don't need to write in a specific way just for generative AI search."

That is a direct primary-source contradiction of a lot of GEO advice, and the honest reconciliation is that Google is describing what's required, not what's advantaged. Its guide also says SEO best practices "continue to be relevant because our generative AI features on Google Search are rooted in our core Search ranking and quality systems." Meanwhile the GEO-bench experiment and the absorption study both show extraction-friendly prose gets used more heavily once a page is retrieved.

Both statements survive if you frame the work correctly: write for a reader, and make sure each section survives being lifted out of context. There's no secret format. There is such a thing as a paragraph that's easier to quote.

Schema sits in the same bucket. The same Google guide says "structured data isn't required for generative AI search, and there's no special schema.org markup you need to add." Search Engine Land notes Google has confirmed structured data gives an advantage in search results and Microsoft's Fabrice Canel has said schema helps LLMs understand content for Copilot — but a December 2024 study found no correlation between schema coverage and citation rates, and there are no controlled experiments here. Aimee Jurenka's framing is right: "Schema markup is infrastructure, not a magic bullet." Valid frontmatter, correct meta lengths and a clean slug keep a page eligible. Hygiene, not a cheat code — the same distinction we drew in our close reading of Google's scaled content abuse policy, where the offence was never the method but volume without value.

How to tell if it's working when the analytics lie

Attribution in AI search is genuinely broken, and you should know that before you run an experiment and misread the result.

Referrer data is incomplete. ChatGPT citation links carry utm_source=chatgpt.com for logged-in users, but free-tier traffic arrives with no referrer and lands in GA4 as Direct. Your dashboard is undercounting by an unknown amount.

Server logs are the only free view of the crawlers. Because AI bots don't execute JavaScript, they're invisible to client-side analytics entirely. Grep your access logs for OAI-SearchBot and ChatGPT-User — a five-minute job that answers the crawler-access question definitively.

Google Search Console now shows AI impressions — and only impressions. Google launched a Search Generative AI Performance report on 3 June 2026, isolating impressions inside AI Overviews, AI Mode and AI Overviews in Discover. The limitations are significant: impressions only, no clicks, CTR, average position or queries, the three surfaces blended and inseparable, and a phased rollout to a subset of sites. A directional signal, not a measurement instrument yet.

And the platforms themselves are unstable. seoClarity tracked ChatGPT citation patterns from 8 February to 27 April 2026 across five markets and found citation volumes fell 86–94% across all of them between February and April. The US zero-citation rate doubled from 28% to 48% in March; Germany's reached 85% by April. Citations then recovered toward pre-March levels in May. The pattern was volatility, not decline.

The practical consequence: any before/after test you run over four weeks is probably measuring platform noise, not your changes. Give a page-craft experiment a quarter at minimum. seoClarity's own read is worth keeping: "AI visibility is becoming less about winning individual links and more about building durable brand authority inside the model's understanding."

What a citation is actually worth

You have watched informational traffic drain into the answer box, so you know the top-line numbers. They're worse than the vibes suggest, and the upside is narrower — but better — than the hype suggests.

The drain is real. SparkToro's analysis of Similarweb's US desktop and mobile-web panel found 68.01% of US Google searches ended without a click in January–April 2026, up from 60.45% in 2024 and roughly 49% in 2019 (browser-based searches only; excludes the Google mobile app). Pew Research Center, tracking 900 US adults' actual browsing across 68,879 Google searches in March 2025, found users clicked a traditional result in 8% of visits when an AI summary appeared versus 15% when it did not — about half — and only 1% clicked a link inside the AI summary itself. Ahrefs' December 2025 update, comparing 300,000 keywords against aggregated Search Console data, found position-1 desktop CTR fell 58% where an AI Overview is present. Ryan Law's summary: "For every 100 clicks you could historically earn for a top-ranking page, Google now keeps 58."

And AI referral traffic is tiny. The same SparkToro analysis found AI tools collectively send less than 1% of all traffic out. Anyone promising you a traffic replacement is selling something.

But the traffic that does arrive converts unusually well. Ahrefs' own first-party data over 30 days found AI search accounted for 0.5% of traffic but 12.1% of signups — roughly 23× the conversion rate of traditional organic. Semrush, studying 500+ high-value digital marketing and SEO topics, concluded that the average AI search visitor is 4.4 times as valuable as the average traditional organic visit on conversion rate, and projects that for those topics AI search may drive more visitors than traditional search by early 2028. Both are single-vertical, self-reported figures; treat them as encouraging rather than universal.

With a counterweight from the same publisher. Ahrefs' study of about 82,000 websites in May–June 2025 found AI visitors are less engaged on average — 4.0 pages per visit versus 5.2 from search, 67.8% bounce versus 63.7% — which they characterise as "information validation or one-and-done behaviour."

One more data point complicates anyone's plan, including this post's. Ahrefs looked at where its own AI search traffic lands: 80% went to free tools (36.45%), product pages (23.12%) and the homepage (20.41%) — not to its blog. Patrick Stox: "We have the best content in the entire SEO industry, but all of our content that aligns with what search engines are telling us to do is driving only a fraction of the traffic from AI search." The synthesis: blog paragraphs win citations; product pages, tools and comparison pages win clicks. Do both, and don't confuse which one you're measuring.

Content type matters more than generic advice admits, and it's industry-specific. DeltaV Digital tracked 21,075 AI engine responses and 25,337 citations over 90 days (14 April – 13 July 2026) across five platforms for eight brands in eight industries, and found listicles earn 61% of AI citations in B2B technology services and 0% in healthcare, where articles take 54%. Their most useful distinction is citation share versus citation rate — citations earned per retrieval. Comparison pages held just 4.1% of share but the highest rate at 1.87, about 45% above their portfolio average. Their illustration: AI engines retrieved WebMD constantly and cited it 0.33 times per read; Mayo Clinic scored 1.79. Retrieval is not citation. The difference is what the passage offers once the page is opened.

Where the widest doors are, per DigitalApplied's intent breakdown: definitional queries average 5.6 citations per AI Overview and how-to queries 5.1, against 3.1 for commercial queries. If you're picking what to write next, definitional and procedural posts on narrow topics have measurably more room.

The short version, in order

1. Fix the floor first. Server-rendered HTML, clean sitemap, no dead internal links, OAI-SearchBot allowed, no stray nosnippet or tight max-snippet, and a check on your Cloudflare bot rules before 15 September 2026.
2. Write one committed sentence per section. Definition, number, comparison or step — under an H2 that matches the literal question, with no preamble above it.
3. Attach a named source and a number to every claim that carries weight. Then verify the link actually resolves and the page actually says it.
4. Publish original data where you have any. Your own logs, your own benchmarks, your own before/after. It is the one thing a model cannot generate for itself.
5. Put a visible last-updated date on every post and honour it. A quarterly refresh pass on 20 posts beats 20 new posts.
6. Measure in quarters, in server logs and in Search Console impressions — not in four-week GA4 experiments.
7. Keep a cadence. Citation accrues to pages that exist, stay accurate and stay maintained.

That last item is where most small blogs fail — not on craft, but on the sixth month. If holding a steady cadence at full depth is the constraint, that's the problem Magic Share was built for: the agent researches your niche, drafts a 2,000–4,000-word post (sized to the topic) with internal links, verifies that 100% of cited URLs resolve, validates frontmatter and metadata against schema rules, and holds the result as a draft until you approve it. Merging the pull request is the publish button, and that's yours.

FAQ

Do I need llms.txt to get cited by ChatGPT or AI Overviews?

No. Ahrefs' May 2026 server-log study of 137,210 domains found 28% publish an llms.txt file and 97% of those files received zero requests that month. Gary Illyes has confirmed Google doesn't support it and has no plans to, and Google's own optimization guide states you don't need to create machine-readable files, AI text files, markup or Markdown to appear in its generative AI features. It costs nothing to ship, but expect nothing from it.

For a deeper, evidence-based look at llms.txt, Google's stance, and practical next steps for publishers, see analysis of llms.txt and its SEO impact.

Do I need to rank in the top 10 to appear in an AI Overview?

Not necessarily. Ahrefs' March 2026 analysis of 863,000 keywords and 4 million AI Overview URLs found 38% of citations came from top-10 pages, 31.2% from positions 11–100, and 31.0% from pages ranking beyond position 100. BrightEdge, measuring across nine industries, put the top-10 share at 16.7% and described positions 21–100 as the sweet spot. Ranking well still helps; it isn't the gate.

Does schema markup get you cited?

The evidence is thin. Google states explicitly that structured data isn't required for its generative AI features and that there's no special schema.org markup to add. One 1,000-AI-Overview analysis found pages with Article and BreadcrumbList schema were cited 2.3× more often, but that's correlational, and a December 2024 study found no correlation between schema coverage and citation rates. Treat schema as infrastructure that keeps a page eligible and machine-readable, not as a lever.

How long before a small blog starts getting cited?

Plan in quarters. The one detailed public case study — an agency's own, with an anonymous bootstrapped client — reported zero citations for two months, first long-tail citations at months 3–4, and consistent cross-platform citations at months 9–10. That's unaudited, but it matches the volatility picture: seoClarity measured ChatGPT citation volumes falling 86–94% across five markets between February and April 2026, then recovering in May. A four-week test can't separate your changes from that noise.

Should I update old posts or publish new ones?

Both, weighted toward updating if your library is small. Seer Interactive's study of 7,683 dated pages and 47,097 citations found 75% of cited pages were updated within the past year, but only 42% were originally published within the past year — 27% of "fresh" cited content was more than two years old at publication. New posts get citation spikes; maintained posts get durable citation. A visible, honest last-updated date is the cheapest thing on this list.

Where this leaves a 20-post blog

The realistic promise is not that you will out-cite Wikipedia or Reddit inside AI answers. Twelve domains take roughly half the citations in AI Overviews, and that is not changing this year.

The realistic promise is the other half — the long-tail, definitional and procedural queries where AI Overviews cite five or six sources, and where the deciding factor is whether your paragraph committed to an answer, named its source, carried a number, and was reachable by a crawler that doesn't run JavaScript. Those are all page-level decisions, all inside your control, and none of them require a budget.

The hard part is doing it on the twentieth post as carefully as on the first, for four quarters, while running everything else. If that's the bottleneck, see Magic Share's plans — your first three posts are free. Or read our other notes on publishing AI-assisted content honestly and go check your robots.txt first. That one's free either way.