AI search optimization — what works and what's snake oil

Most of what’s sold as AI search optimization — GEO, AEO, “AI visibility” — is ordinary SEO repackaged at a premium: typical retainers run $1,500–$5,000 a month for small businesses, above the $1,000–$2,500 average for the SEO work they mostly repeat. A few levers measurably move AI visibility: ranking in the indexes the engines read, verifiable substance in your content, genuine freshness, and off-site brand mentions. Much of the rest — llms.txt files, schema-for-citations, guaranteed AI rankings — is contradicted by public measurement, some of it by the vendors’ own tests.

One disclosure makes this page different from everything else ranking for these questions: we don’t sell GEO services, and we have no plans to. We’re a paid-media agency; the measured levers below are baked into how we build pages anyway, not sold as a separate line item. This is the briefing we’d give a client who asked whether to sign that $3,000-a-month AI-visibility retainer.

How AI search actually picks its sources

Every AI answer engine works the same way at the spine: it runs searches against an index, reads the top results, and synthesizes an answer with citations. Which index is the detail that matters — and it’s why “GEO without SEO” isn’t a real thing:

Engine figures: seoClarity’s AI Overviews overlap research, FelloAI’s Perplexity citation analysis, and Profound’s Claude–Brave study, all checked July 11, 2026.

Google says it plainly in its official documentation (checked July 11, 2026): no special files, no special markup, no AI-specific optimizations — its AI features are “rooted in our core Search ranking and quality systems.” If someone can’t get you ranked in ordinary search, they cannot get you cited by the machines that read it. That single sentence disqualifies most GEO pitches.

What measurably works

Four levers have real data behind them — every one a thing good marketing already does:

Rank in the underlying indexes. See the table above. The one nuance worth money: since Google’s January 2026 update, AI Overview citations spread far beyond the top-10 (the share drawn from top-10 results fell from 76% to 38% — Ahrefs, 863,000 SERPs, January 2026) because queries get decomposed into sub-questions — so a page section that precisely answers a narrow sub-question can get cited without ranking for the head term. That rewards structure, not tricks.

Verifiable substance. The only peer-reviewed result in this field (Princeton’s GEO paper, KDD 2024) measured +30–40% visibility for content carrying statistics, quotations, and cited sources. The honest caveats: it was a lab setup, and the follow-up replication found that rewrite-only tricks — tone, phrasing, keyword variants — fail or backfire. Only adding real, checkable substance survives scrutiny. Dated numbers with named sources are the tactic; this page is the demonstration.

Freshness — the honest version. Across ~17 million citations, AI-cited content averages 25.7% fresher than what wins ordinary search (Ahrefs, July 2025). But the average cited page is still 2.9 years old, and the study’s authors warn that low-effort re-stamping does nothing. Hold that against the “content updated within 30 days gets 3.2x more citations” claim circulating on vendor sites — a number whose sourcing is two AI-visibility vendors citing each other, with no methodology published at either end.

Off-site brand mentions. In the largest correlation study available (75,000 brands, Ahrefs), branded web mentions correlate 0.664 with AI visibility and YouTube mentions 0.737 — while backlinks, the currency of classic SEO, correlate just 0.218. AI assistants recommend brands they’ve read about across reviews, comparisons, Reddit, and editorial coverage. That’s earned media — PR with a new scoreboard — and it’s slow, real work no file upload replaces.

The snake-oil ledger

Every claim below is being sold right now — all examples checked July 11, 2026.

The llms.txt file. Sold via guides and generators as an AI-visibility play. Measurement: SE Ranking analyzed ~300,000 domains and found no correlation between having the file and AI citations — removing it as a variable improved their model’s accuracy. Google’s docs state outright that no machine-readable AI files are needed, and crawl-log studies show almost no AI bot ever requests it. If an agency’s deliverables list includes llms.txt, you’re paying for a file nothing reads.

“Schema markup increases AI citations.” The cluster’s biggest unsupported claim, and its receipts are remarkable. The widely circulated “GPT accuracy improves from 16% to 54% with structured data” traces to a 2023 benchmark of GPT-4 answering database-style questions over an internal enterprise knowledge graph — nothing to do with schema.org markup on webpages. A “55% AI visibility boost” case study measures a score generated by the schema-selling tool itself; the same post concedes no referral traffic materialized after 30 days. Meanwhile a GEO-tool vendor’s own three-month, 319-prompt experiment found six of seven AI platforms couldn’t even fetch or interpret schema markup, and Google’s documentation says no special structured data is needed for AI features. Schema earns rich results in classic search — that’s the honest reason to keep it. Citations, no.

Guaranteed citations, “AI submission,” and 45-day miracles. “0% to 72% AI visibility in 45 days” is a vendor-defined score over a vendor-chosen prompt set — unfalsifiable by design. The structural problem with all guarantee-shaped promises: only about 30% of brands stay visible across back-to-back runs of the same AI query (NAV43 tracking data, 2026). And there is no submission mechanism to any AI engine — a warning that, to their credit, some AI-native agencies now publish themselves. Anyone selling “proprietary submission” is selling a door that doesn’t exist.

Chunk optimization at 2–4x. Writing self-contained, answer-first sections is legitimate — it’s how retrieval works, and we do it on every page. The “2–4x citation improvement” multiplier attached to it by vendors has no study, no dataset, and no named source anywhere we could find.

The “best GEO agency” lists. The top-ranking “top GEO agencies” page ranks its own author #1, with a self-assigned five-star “AI Visibility Score” of its own design. Every list on that results page features or favors its author. There is currently no independent evaluation of GEO providers anywhere on page one — worth knowing before treating those rankings as research.

What it costs — and what a fair price buys

The going rates (pricing census, July 11, 2026): GEO agency retainers run $1,500–$5,000/month for small businesses and $5,000–$25,000 mid-market; one-off “GEO audits” $500–$2,500; AI-visibility tracking tools from $29/month (entry tiers, ~15 prompts) to $499+/month. For contrast, traditional SEO averages $1,000–$2,500/month (Backlinko survey of 300+ practitioners, December 2025) — the premium is for the acronym.

A fair engagement does exist. If the deliverables are answer-first content with real substance, digital PR that earns third-party mentions, technical hygiene, and visibility reporting that’s honest about being a sample of non-deterministic answers — that’s real work under a new label, and worth normal content-and-PR money. The walk-away signs: llms.txt on the deliverables list, schema sold as a citation lever, any guarantee, any “submission,” and case studies measured in scores the seller invented.

What just changed (June–July 2026)

Three developments this ledger’s targets haven’t caught up with. Google Search Console now reports AI visibility first-party — a generative-AI performance report launched June 3, showing impressions in AI Overviews and AI Mode (no clicks, no queries yet, rolling out gradually). Before paying for third-party “AI rank tracking,” check what Google now shows you free. Ads are inside the answers: ChatGPT ads went live in the US in February and reached the UK by June (per OpenAI’s announcements, with the conversion-optimized rollout reported by Search Engine Land) — paid placement in the answer box changes the math on chasing organic citations, and it’s a channel bought like any other paid platform, not through a GEO retainer. And Cloudflare set a September 15 deadline for AI companies to separate their search crawlers from training crawlers or face default blocking across a large slice of the web — the “maximize AI crawler access” advice in most GEO playbooks is about to collide with the infrastructure layer.

The verdict

AI search is worth your attention and almost never worth a panic budget. The measured reality: AI assistants send roughly 1% of referral traffic for most businesses — but early tracking studies report those visitors converting anywhere from 4.4x to 23x better than average, because they arrive pre-sold by the answer. The winning move is unglamorous: be excellent in ordinary search, put checkable facts in your content, keep it genuinely current, and earn third-party mentions. That’s not a new discipline with a new invoice. It’s the standard we’d hold any marketing spend to — including ours: if you can’t see the mechanism and the measurement, don’t buy the acronym.

Common questions

Do I need GEO for my small business?

Almost certainly not as a paid retainer. AI assistants drive under 1% of referral traffic for most businesses (Ahrefs), and the levers that matter — ranking in search, substantive content, reviews and mentions — are things you’re already buying or building. Do the free hygiene: check your site isn’t blocking AI crawlers, keep your Google Business-class listings and reviews current, and write pages that answer questions directly.

How do I get my business mentioned in ChatGPT?

There’s no submission form — assistants recommend brands they find in their index and training data. The convergent advice, consistent with the mention-correlation data: current reviews on the platforms your buyers use, a presence on Bing (ChatGPT’s index was seeded from it), and third-party coverage — comparisons, directories, editorial mentions. Brand mentions correlate with AI visibility at 0.66–0.74; backlinks at just 0.218.

Is SEO dead because of AI?

No — it’s the entry ticket. Every major engine cites what its underlying index ranks (Gemini literally runs on Google Search; Claude’s citations overlap Brave’s results 86.7%). What did change is the economics of clicks: position-1 organic CTR falls ~58% when an AI Overview appears (Ahrefs, December 2025), so the same rankings return fewer, better visitors. Content built to be the cited source — direct answers, real numbers — is how you stay in the game.

Can you actually track AI rankings?

Not the way rank trackers track Google. AI answers are non-deterministic — roughly 30% of brands stay visible across back-to-back runs of the same query — so honest tools report visibility percentages across repeated sampled prompts, not positions. Since June 2026, Google Search Console shows AI Overview impressions first-party and free. Distrust any report with a “rank” and no confidence interval.

Does schema markup help AI visibility?

Keep schema for what it demonstrably does: rich results and disambiguation in classic search. As an AI-citation lever, the evidence is against it — a GEO vendor’s own 319-prompt, three-month test found six of seven AI platforms couldn’t fetch or interpret it, and Google’s docs say no special structured data is needed for AI features. The viral “16% to 54% accuracy” stat comes from a 2023 enterprise database benchmark, not web search.

What should AI search help cost?

If it’s folded into competent SEO/content work at normal rates ($1,000–$2,500/month is the surveyed SEO average), you’re paying fairly for real work. Stand-alone GEO retainers at $1,500–$5,000/month deserve the deliverables test: substance, PR, hygiene, honest sampling-based reporting — fine; llms.txt, schema-for-AI, submissions, or guarantees — walk. A $500–$2,500 “GEO audit” mostly re-sells a technical SEO audit with new vocabulary.

Every claim on this page was checked on July 11, 2026 — the measurement studies, Google’s documentation, the vendor pages making the claims, and their sources’ sources. This is the fastest-moving topic we cover: we re-verify it quarterly against a standing list of facts most likely to shift, and mark real updates with a visible date. If a claim here has been overtaken by better measurement, we’ll change it — that’s the difference we’re selling.