775 AI Citations Measured: Your Own Content Earns 14%, Third-Party Listicles Take 85.8%

On September 14, Search Engine Land published consultant Zeeshan Yaseen's write-up of two back-to-back GEO experiments: two brands, six AI platforms, every query run and logged by hand, 775 citation events in total. The number worth keeping is not the total but the breakdown. Across 437 source mentions in the second experiment, the brand's own listicle on its own domain accounted for 14.0%, third-party listicles for 85.8%, and press releases for 0.2%. The owned page also took 18 days to earn its first citation, slower than every third-party placement.

That lands squarely on the GEO industry's core pitch. Over the past 18 months, AI visibility tooling grew into its own category selling content-as-asset: rewrite your site so models parse it, bolt on schema, publish an llms.txt, watch a dashboard score. Google's own Search Central documentation says something much flatter — a page only needs to be indexed and snippet-eligible to be eligible as a supporting link in AI Overviews or AI Mode, and "you don't need to create new machine readable files, AI text files, or markup". Search Engine Land's local reporting this year separately found that earning an AI recommendation can be up to 30x harder than ranking in Google. Supply keeps shipping tools; the demand-side ceiling is structural.

The category is splitting too. Published pricing: Otterly.ai runs $29 / $189 / $489 per month for 15 / 100 / 400 tracked prompts; Peec AI starts at $95; Profound charges $99 for a ChatGPT-only Starter and $399 for Growth, both annual-billed; Ahrefs Brand Radar publishes $398 and $699 tiers (2026 pricing roundup). Feature gaps are narrowing while prices spread twentyfold, which means the fight is over becoming the default procurement line item. On the same day, Google was caught testing an AI contribution pilot that pays publishers for content used in AI Mode, AI Overviews and Gemini — effectively pricing the act of being a cited source.

For a small business or freelance studio the read is blunt: if 85.8% of your citations come from pages you do not control, "write better site content" is not priority one, it is the entry requirement. What follows is how those 775 citations distributed, which moves measurably shifted the numbers, and how a five-person team runs its own version without a subscription.

What happened, with the numbers

Experiment one was the consultant test: an established client, several months, thousands of dollars, 15 commercial-intent keywords across ChatGPT, Claude, Gemini and Perplexity, every query run manually and repeated with and without a VPN. The brand ended up appearing for 10 to 12 of the 15 keywords, peaking at 37.01% keyword presence on April 29. Citations by platform: ChatGPT 148, Claude 96, Gemini 87, Perplexity 64. A single comprehensive listicle on Indeed SEO generated 190 citations on its own, more than every other source combined. Listicles drove 72.4% of citations, PR 24.1%.

Experiment two was the cold start: a SaaS link building agency with zero measurable presence at kickoff. Over 30 days (May 30 to June 28) across six platforms it produced 298 appearances from nothing — Gemini 104, Google AI Mode 95, Claude 59, ChatGPT 32, Grok 4, Perplexity 4. Google surfaces took roughly two-thirds, the inverse of experiment one.

Three mechanical findings matter more than the totals. Concentration is extreme: three sources accounted for 342 of 437 source mentions, about 78%. Building the outreach list from observed citations rather than domain rating raised the hit rate — Indie Hackers climbed from 44 mentions to 146 after placement (+232%), Bruce Jones SEO from 26 to 69 (+165%), TechBullion from 15 to 37 (+147%) — though RankTracker and HR.com moved the other way. And comparative content beat promotional content: on June 23 the owned listicle was rewritten to include leading competitors rather than featuring the brand alone, and mentions of that source rose from 4 to 49, a 12.25-fold increase.

The revenue side is just as instructive. During the window, 18.5% of new users arrived via referral and 3.25% via the AI Assistant channel in GA4's default channel group — together just over a fifth of all new users. The deal that closed came from Perplexity, the lowest-volume platform at four appearances.

What each reader should do this week

Brand owners and SMB operators

• Spend two hours running the 10 sentences a real buyer would type through ChatGPT, Gemini and Google AI Mode, and write down which domains get cited. That list is your budget allocation plan.
• Cut the blog budget in half and move it toward earning placement in third-party listicles. The measured split is 14% owned versus 85.8% earned.
• Budget for maintenance, not a project: roughly half of all sources stopped being cited within 30 days, and one placement fell from 29 mentions to 11 week over week.

Marketing and SEO practitioners

• Treat your peer set as a lever. Removing recognized names preceded a decline within days; adding named competitors preceded a 12.25x rise. Models appear to weigh the entities around you.
• Put the answer in the first 100 words and add a key-takeaway block at the top — the highest-yield on-page change in the tests. Keep FAQs expanded rather than hidden in accordions.
• Give high-value commercial phrases dedicated assets. The brand ranked for "Best LLM SEO Consultant" but barely appeared for the near-synonymous "Best AI SEO Consultant".
• Extend your measurement window: time-to-citation ranged from 1 to 18 days, so a single check one week after publication misleads.

Developers and agencies

• Productize citation prospecting: a scheduled script that runs 15 prompts weekly, extracts cited domains and renders a trend chart. Low difficulty, and clients will not build it.
• In the client's GA4, isolate the AI Assistant channel in an exploration with a referral-domain dimension, keeping "cited" and "clicked" separate.
• Add FAQPage and Service structured data to comparison pages so the extraction layer guesses less.

Tool comparison

The measurement delta between these four is usually smaller than the delta from spending the same money on one well-chosen third-party placement. Tools solve knowing, not moving.

What nobody selling this will tell you

• Citations are not revenue. Indie Hackers led on citation volume while its referral traffic stayed flat; TechBullion produced far fewer citations but took sessions from 1 to 64. A dashboard visibility score as your KPI keeps funding sources that never send a customer.
• AI-specific files have no official backing. Google's documentation states no new machine-readable file or markup is required to appear in AI features. Paying for one buys reassurance, not eligibility.
• The two experiments contradict each other, and the author says so. Round one recommended publishing listicles on your own site; round two downgraded owned content to a foundation rather than a growth lever. Anyone selling a settled 2026 playbook is more confident and less evidenced than the person who ran the tests.
• The sample is two brands, English-language queries. The source pool for other languages differs, so the platform split does not transfer.

The no-subscription version for small teams

A spreadsheet and one script, at roughly 90 minutes a week:

• Step 1 — Build the prompt set. Write 15 sentences real buyers would type, covering recommendation, comparison, pricing and shortlist intents.
• Step 2 — Take a baseline. Run each prompt once on ChatGPT, Gemini, Google AI Mode and Perplexity, recording three fields: were you mentioned, which domains were cited, where you appeared in the ordering.
• Step 3 — Rank your channels. List the 10 most frequently cited domains as your pitch and inclusion list, replacing domain-rating-based prospecting.
• Step 4 — Publish one comparative page. Include competitors honestly, answer in the first 100 words, FAQs expanded, question-form headings. This is the foundation, not the engine.
• Step 5 — Track on two rails. Visibility from your weekly manual log, business from GA4's AI Assistant channel and referral domains. When they diverge, trust the second.

Self-check:

• ☐ 15-prompt question set written
• ☐ One baseline round completed with cited domains recorded
• ☐ Top 10 cited domains ranked into an outreach list
• ☐ Comparative page includes at least 3 competitors
• ☐ GA4 isolates the AI Assistant channel
• ☐ A 30-day decay check is scheduled

FAQ

Should I still write content on my own site?

Yes, but reposition it. Owned content drove 14% of citations and took 18 days to be cited at all. Its job is to hold up when a model verifies you. Exposure comes from third-party pages, so the spend ratio should flip from mostly-owned to mostly-earned.

Are llms.txt and AI-specific schema worth doing?

Google's documentation is explicit: a page only needs to be indexed and snippet-eligible to be eligible for AI features, with no additional machine-readable file or markup required. Established structured data like FAQPage and Product still earns its keep, but it is not an AI visibility switch.

How do I monitor mentions without paying for a tool?

A manual prompt set and a spreadsheet are enough. The gap that matters is not measurement precision but whether you convert measurement into an outreach list. GA4's default channel group already includes AI Assistant, so sessions from ChatGPT, Gemini and Claude are visible at no cost.

Does this data transfer to non-English markets?

The platform split does not — this is a two-brand, English-language sample. The method does: measure who gets cited, go earn placement there, budget for decay, and replace promotional pages with comparative ones.

My take

The consensus says GEO is a content problem. I think it is a distribution problem that collapses back into relationships. The two variables with the most explanatory power across 775 citations were whose page you appear on and who you appear next to. Neither is a writing skill; both are media relations and industry network. So I expect that within 12 to 18 months pure measurement tools get compressed into a thin price band, or absorbed as a feature inside suites like Ahrefs and Semrush. Vendors sell a dashboard; buyers want someone to get their name onto the list.

For ScriptWalker the productization is specific: package the workflow above as a recurring "AI citation source prospecting plus quarterly maintenance" retainer. Month one delivers the baseline report and a ranked top-10 cited-domain list; each month after delivers a change log plus one refreshed comparative page. Low technical barrier, highly automatable, and renewal is structurally justified because sources decay. It sells the part clients cannot do themselves and tools cannot do at all.

Sources

• Primary | Two GEO experiments challenge conventional AI visibility advice (Search Engine Land, 2026-09-14)
• Primary | AI features and your website (Google Search Central documentation)
• Primary | Default channel group (Google Analytics Help)
• Primary | Google tests paying publishers for using its content in AI Mode, AI Overviews and Gemini (2026-09-14)
• Third-party | AI local visibility is up to 30x harder than ranking in Google: Report
• Third-party | AI Visibility Tool Pricing Compared (2026)
• Third-party | Daily Search Forum Recap: September 14, 2026 (Search Engine Roundtable)

Want help running this?

If you want the citation prospecting, the GA4 channel split and the comparative page done in one pass, take 30 minutes with us first so we can tell you whether your category justifies the spend.

• Email: [email protected]
• Phone: 0916-224-047
• LINE: @ufv9089p