Is AI SEO worth paying for? What the studies actually found.
The going rate is $1,500 to $12,000 a month, and most of what it buys is technical — schema markup, llms.txt, crawler configuration. Google published guidance in May 2026 explicitly saying most of it isn't needed. The strongest experimental test found no measurable effect. A correlational study found a small one. Both are real, they disagree, and almost nobody selling this mentions the first. Here's what the evidence says on each side, and the part where I correct my own copy.
For most local businesses, no — not as a separate product at $3,000 to $6,000 a month. Google published official guidance in May 2026 saying there is no special schema.org markup needed for AI search, that llms.txt gets no special treatment, and that content chunking isn't required — which is most of what these retainers bill for. A controlled test of 1,885 pages that added schema found no measurable change in AI citations either, and none of the five major engines read schema when they fetch a page live. A correlational study found a small positive association, so it isn't settled — but nothing in either result justifies that price. What the evidence does support is off-site presence, cited statistics and freshness, and those are cheaper.
What it costs, and what the price buys
Published pricing for AEO and GEO services — the two names this gets sold under — runs from about $1,500 a month at entry level to $12,000 or more for a full program. Mid-size businesses are typically told to budget $3,000 to $6,000 a month.[1]
Here's what that covers, quoted from a provider's own pricing page: "direct answer blocks, FAQ schema, and Speakable markup so that LLMs can extract factual answers." And on why it costs more than regular SEO: "the technical layer requires specialized schema markup, robots.txt configuration for AI-specific crawlers, and IndexNow protocol implementation."[1]
Another agency describes its GEO layer as "llms.txt, AEO schema, citation engineering."[2] The pattern is consistent: the premium is justified by technical work on your own website.
And here's the sales method, published in a guide teaching agencies how to sell it: "The most effective approach is a live demo. Ask the client 'Do you know what ChatGPT says about your brand?' and then show them in real time. When they see competitors getting mentioned and they don't, the conversation shifts from 'why should I care?' to 'how fast can we start?'"[3]
The same guide notes that for the agency, "the tooling costs are low (starting at €49/mo)."[3]
I want to be careful here. A fear demo isn't fraud, and the gap it demonstrates is real — ChatGPT genuinely does recommend a tiny fraction of local businesses. The question isn't whether the problem exists. It's whether the thing being sold fixes it.
Google answered this in May, in writing
Here's the part almost nobody selling AI SEO will mention, and it's the highest-authority source available on the subject.
On May 15, 2026, Google published its first official guidance on optimizing for generative AI features — a Search Central document written by John Mueller.[13] It opens by asking whether SEO still matters for AI search and answers itself: "In short, yes!" AI Overviews and AI Mode are, in Google's own words, "rooted in our core Search ranking and quality systems" — the same index, the same crawl, the same quality signals as ordinary results.[13]
The document includes a section called "Mythbusting generative AI search", which names the tactics you can skip. It reads like a list of what's currently being invoiced.[13]
| Sold as necessary | What Google says |
|---|---|
| Special AI schema markup | "Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add." Google adds that it's still worth using for ordinary rich results.[13] |
| llms.txt files | Google may crawl one like any other file, but "this doesn't mean that the file is treated in a special way." Mueller has stated no AI system uses it.[13] |
| Content chunking for parsers | "No requirement to break your content into tiny pieces" and "no ideal page length." Structure it for readers.[13] |
| AI-specific rewrites and Markdown copies of pages | Not needed. The mythbusting section names these directly.[13] |
And a year earlier, at WordCamp US in August 2025, Google Search's Danny Sullivan had already put it more bluntly from the stage: "Good SEO is good GEO or AEO or AI SEO or LLM SEO or even LMNOPEO."[14]
The part with teeth
Since May 15, 2026, manipulating AI answers is spam by Google policy.[15] No bought mentions, no manipulation, no shortcut discipline. As one analysis put it, Google published the door and the lock on the same day — which suggests the company saw where the vendor market was heading and got ahead of it in writing.
Google also published separate guidance on how to evaluate SEO and AI tools before paying for them.[15] That is not a coincidence either.
⚠ The caveat that keeps this honest, and it matters. Google's guide speaks for Google Search, which powers AI Overviews and AI Mode. It says nothing about ChatGPT, Perplexity, standalone Gemini or Copilot — each has its own retrieval sources and its own habits.[16] So "still SEO" is authoritative for Google's surfaces and only indicative elsewhere. The next article covers how differently those engines behave.
The experiment that found nothing
This is the strongest test anyone has run on the technical layer, and it's the one that rarely appears in sales material.
Ahrefs analyzed 6 million URLs and tracked 1,885 pages that added JSON-LD structured data between August 2025 and March 2026. Adding it produced no measurable change in citations from Google AI Overviews, Google AI Mode, or ChatGPT in the 30 days afterwards.[4]
They also tested whether the engines read schema at all. When ChatGPT, Claude, Perplexity, Gemini and Google AI Mode were asked to fetch a page live, none of the five read the schema — all extracted only the visible HTML.[4]
That second finding is the one worth sitting with. If an engine reading your page in real time doesn't parse your structured data, then markup written specifically to be read by that engine isn't doing the job it was sold to do.
There's a plausible technical reason. When these systems generate an answer they work from processed index summaries and grounding passages, not a direct feed of raw JSON-LD — and one experiment found AI crawlers tokenize JSON-LD as plain text rather than parsing its semantic structure.[18] There is no separate "AI schema pipeline" for markup to travel down.
The statistic the industry quotes, explained
You'll see it claimed that AI-cited pages are far more likely to have schema. That's true — Ahrefs found cited pages were nearly 3x more likely to already carry JSON-LD.[4]
But Ahrefs attributes that to better-maintained sites, not to the markup itself.[4] Sites that bother with schema also tend to bother with everything else — fresh content, clean structure, real coverage. The schema is a symptom of a well-run site, not the cause of the citation.
Which is exactly why the before-and-after test matters more than the correlation. They added markup to 1,885 real pages and watched. Nothing moved.
The study that found something
I'd be doing the same thing I'm criticizing if I stopped there, because the evidence isn't unanimous.
AirOps' 2026 State of AI Search report found that pages with three or more schema types show a 13% higher citation likelihood, and that 68.7% of ChatGPT-cited pages follow logical heading hierarchies.[5] Another analysis argues local businesses with proper LocalBusiness markup get cited more on geo-targeted queries.[6]
A third study lands between the two and is probably the most useful of the three. Trakkr examined 28,000+ citation appearances across 950 domains and found that while 68% of AI-cited pages carry structured data — double the web average — schema types don't predict citation volume.[17] Cited pages have schema; having schema doesn't get you cited. Content quality and structure did the predicting.
| The experiment | The correlation | |
|---|---|---|
| What was done | 1,885 pages had JSON-LD added, then were watched for 30 days[4] | Cited and uncited pages compared for schema presence[5] |
| What was found | No measurable change in citations | 13% higher citation likelihood with 3+ schema types |
| What it can show | Whether adding it causes anything | Whether the two things appear together |
| What it can't | Effects beyond 30 days, or on other engines | Which one caused which |
Both are real results. They're answering slightly different questions. A before-and-after test is stronger evidence of cause than a comparison of two groups — but 30 days is a short window, and no single study settles a field this new.
The honest conclusion, and it's narrower than either camp would like: schema is cheap, it's a one-time job, it still earns rich results in normal search, and it does no harm. Keep it. What the evidence does not support is paying thousands a month for it as an AI-citation strategy, which is exactly what the pricing above is built on.
llms.txt, and other things scored honestly
In May 2026 Cyrus Shepard of Zyppy published something the field badly needed: a meta-analysis synthesizing 54 experiments, patents and case studies into a single scored ranking of 23 factors — the first attempt to weight this advice by evidence strength rather than opinion.[7]
| Sold as | What the evidence says |
|---|---|
| llms.txt — a file telling AI crawlers how to read your site | Zyppy scored it 2.0 out of 10 — no credible evidence it influences citations.[7] A separate April 2026 test found minimal impact.[8] And 28% of domains with traffic already publish one,[4] so it isn't even a differentiator. |
| Press releases for AI visibility | Across 3,600 prompts in 10 industries, press releases captured a negligible fraction of AI news citations.[9] |
| Ranking #1 on Google as the route in | Only 38% of AI Overview citations come from pages in Google's top 10 — down from 76% in mid-2025.[10] |
Zyppy's own framing is worth quoting, because it's the standard the whole field should be held to: "nearly every figure is a study-stated correlation, not proven causation. A high correlation between brand mentions and citations does not prove that adding mentions causes citations; it is equally possible that already-strong brands earn both."[7]
Where my own copy overstated it
I sell a $600 monthly plan. One of its bullet points says "structured data kept clean for AI assistants."
That wording implies more than the evidence supports, and I wrote it before I'd read the Ahrefs test. Keeping structured data clean is genuinely part of what I do, it helps in ordinary search, and I'm not removing the work. But presenting it as an AI-visibility feature was leaning on an assumption I hadn't checked.
I'm not writing this to look honest. I'm writing it because the whole argument of this article is that you should ask what the evidence says before paying for something — and it would be worthless coming from someone unwilling to apply it to his own price list. If you're on that plan and this bothers you, call me and we'll talk about it.
The work I'd defend on that plan is the same as ever: the Google Business Profile that decides where you sit on Google Maps, reviews, citations, local SEO, and content that cites its sources. Which, as it happens, is what the evidence actually supports.
So what is worth doing?
The same meta-analysis that scored llms.txt at 2 out of 10 points somewhere else entirely. So does a controlled academic study from Princeton, Georgia Tech and IIT Delhi presented at KDD 2024, which tested generative-engine optimization methods directly.[11]
| What to do | What the evidence found |
|---|---|
| Put real, cited statistics in your content | The single strongest lever measured — a 41% visibility lift. Optimizing content for generative engines overall lifted visibility by up to 40%.[11] |
| Quote credible sources by name | Helped further on top of the statistics effect.[11] |
| Get mentioned on sites that aren't yours | 82% of AI citations come from earned media; brands are 6.5x more likely to be cited through third-party pages than their own.[12] |
| Keep pages updated | Pages not refreshed quarterly are 3x more likely to lose citations over time.[5] |
| Never keyword-stuff | The same controlled study found it actively lowered AI visibility.[11] |
Read that list again and notice what it isn't. There's no proprietary tooling on it, no monthly platform fee, and nothing that needs a specialist. It's writing accurately, citing sources and being present in the world — which is what good local SEO already looked like back when Google Maps was the only thing worth ranking in.
Where AI actually looks covers the off-site half in detail, and the last piece in this series is the practical version with a way to measure it.
Questions people ask about this
Does schema markup help you get cited by AI?
The evidence is mixed and the strongest test says no. Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026 and found no measurable change in citations from Google AI Overviews, AI Mode or ChatGPT in the following 30 days, and none of the five major engines read schema when fetching a page live. AirOps separately found pages with three or more schema types show a 13% higher citation likelihood, which is a correlation rather than a controlled test. Schema is still worth having for ordinary search results, but neither result supports paying a monthly premium for it as an AI strategy.
What does Google itself say about optimizing for AI search?
On May 15, 2026 Google published its first official guidance on the subject, written by John Mueller. It states that AI Overviews and AI Mode are rooted in Google's core Search ranking and quality systems, answers "is SEO still relevant for generative AI search?" with "in short, yes!", and includes a mythbusting section stating that structured data is not required, that there is no special schema.org markup to add, that llms.txt receives no special treatment, and that there is no requirement to break content into small pieces. Google notes the guidance covers its own surfaces; it does not speak for ChatGPT or Perplexity.
Does llms.txt do anything?
There is no credible evidence that it does. Zyppy's 2026 meta-analysis of 54 studies scored llms.txt at 2.0 out of 10, and an April 2026 test found minimal impact on citation rates. Around 28% of domains with traffic already publish one, so it is not a differentiator either. It costs almost nothing to add, so there is no harm in it, but it should not be a line item on an invoice.
How much does AI SEO cost?
Published pricing runs from about $1,500 per month at entry level to $12,000 or more per month for a full program, with mid-size businesses typically advised to budget $3,000 to $6,000 per month. An agency-facing guide published in 2026 states the tooling required starts at around €49 per month.
What actually increases the chance of being cited by AI?
In controlled academic testing presented at KDD 2024, adding statistics to content was the strongest single lever measured, producing a 41% visibility lift, with citing credible sources helping further and keyword stuffing actively reducing visibility. Separately, 82% of AI citations come from earned media rather than a brand's own pages, and brands are 6.5 times more likely to be cited through third-party sources than through their own domain.
Should a small local business buy an AI SEO package?
For most, no. The technical layer that justifies the premium pricing is where the evidence is weakest, and the practices that are supported — accurate content with cited sources, presence on third-party sites, and regular updates — are part of ordinary local SEO rather than a separate product.
- Fuel Online, AI SEO Pricing 2026: Cost of AEO, GEO & AI Search Strategy, April 2026 — published monthly pricing bands and the stated technical scope those retainers cover.
- Agency service descriptions collected in industry roundups, 2026 — GEO layer described as llms.txt, AEO schema and citation engineering.
- LLM Pulse, GEO Agency Guide: How to Offer AI Search Optimization Services in 2026, August 2026 — the recommended live-demo sales approach and stated agency tooling costs.
- Ahrefs, 2026 — analysis of 6 million URLs; 1,885 pages tracked after adding JSON-LD between August 2025 and March 2026; live-fetch testing across ChatGPT, Claude, Perplexity, Gemini and Google AI Mode; llms.txt adoption across 137,210 domains with traffic.
- AirOps, 2026 State of AI Search — citation likelihood by number of schema types, heading-hierarchy share among ChatGPT-cited pages, third-party origin of brand mentions, and citation decay for pages not updated quarterly.
- Industry analysis, 2026 — argument that LocalBusiness schema improves citation on geo-targeted AI queries. Included as the counter-position; it is an assertion rather than a controlled test.
- Cyrus Shepard, Zyppy, May 2026 — meta-analysis synthesizing 54 experiments, patents and case studies into a scored ranking of 23 AI citation factors, including the llms.txt score and the correlation-not-causation framing.
- SE Ranking, April 2026 — test finding minimal impact of llms.txt on citation rates.
- BuzzStream and Citation Labs, 2026 — analysis across 3,600 prompts and 10 industries finding press releases captured a negligible fraction of AI news citations.
- Ahrefs, March 2026 — share of AI Overview citations originating from pages ranking in Google's top 10, and the decline from mid-2025.
- Aggarwal et al., presented at KDD 2024 (Princeton, Georgia Tech and IIT Delhi) — controlled testing of generative engine optimization methods; statistics measured as the strongest single lever at a 41% visibility lift, and keyword stuffing measured as reducing visibility.
- Muck Rack, analysis of over one million AI-cited links, and AirOps — share of AI citations from earned media and relative likelihood of citation through third-party sources versus owned domains.
- Google, Google's Guide to Optimizing for Generative AI Features on Google Search, Search Central, May 15, 2026 (authored by John Mueller) — including the "Mythbusting generative AI search" section covering structured data, llms.txt, content chunking and AI-specific rewrites.
- Danny Sullivan, Google Search, keynote at WordCamp US, August 28, 2025.
- Google Search Central, May 2026 — spam policy position on manipulating AI answers, and separate published guidance on evaluating SEO and AI tools before purchase.
- Industry analysis of Google's May 2026 guide, 2026 — noting that the guidance covers Google Search surfaces only and does not speak for ChatGPT, Perplexity, standalone Gemini or Copilot.
- Trakkr, 2026 — analysis of 28,000+ citation appearances across 950 domains; share of AI-cited pages carrying structured data versus the web average, and whether schema types predict citation volume.
- Williams-Cook, M., 2026 — schema markup experiment indicating AI crawlers tokenize JSON-LD as plain text rather than parsing semantic structure; with commentary on grounding passages versus raw structured data.
Before you pay a monthly fee for this, let me check where you stand.
Send me your business name. I'll run the prompts a real customer would type, tell you honestly whether you show up, and check your Google listing while I'm in there. No demo built to frighten you, and no plan attached to the answer.
