How Do I Tell AI Search Pitch Is Real

by | Aug 9, 2026 | How Do I Tell

How Do I Tell If an AI Search Pitch Is Real?

Ask the vendor what result would prove them wrong.

That single question sorts most AI search pitches in about thirty seconds. If a claim is built so that no observation could ever falsify it — no traffic number, no log entry, no citation count — it isn't a strategy. It's a subscription.

Below is a working list of claims currently circulating in AI search sales pitches, sorted by how badly they fail. We've deliberately left vendor names out. The claims travel faster than any one company, and several are now embedded in template decks that get resold up and down the industry.

The pitch usually opens the same way

Some version of: "Your customers are asking AI search engines like ChatGPT, Claude, and Google AI who to hire, where to go, and what to buy. LLMs evaluate businesses on structured data, llms.txt files, citation patterns, and content comprehension — different signals than Google SEO."

The premise is real. The mechanism is mostly invented. And the phrase "different signals than Google SEO" is doing the heaviest lifting, because it's what justifies selling you a second product.

Claims that are flatly false

"We'll get your business into ChatGPT's training data." There is no submission mechanism, and training corpora are frozen at a cutoff date. The channel that actually matters is live retrieval — what an engine fetches at the moment someone asks a question. Anyone selling training-data placement is selling access to a door that does not exist.

"Submit your site to the AI engines." No such endpoint exists for OpenAI, Anthropic, Google, or Perplexity. This is the 1999 "submit to 500 search engines" pitch with new logos.

"You need FAQ and HowTo schema to get cited by AI." Google deprecated both rich result types. Its own AI search guidance states that structured data is not required for generative AI features and there is no special schema.org markup to add. Schema still earns its keep — it disambiguates your entity and it's the basis for other rich results — but no AI engine ranks you on it.

"Your content needs to be chunked into AI-digestible pieces." Google's published guidance addresses this directly: its systems understand multiple topics on a single page and surface the relevant section per query. A well-structured long article outperforms one carved up by formula.

"We guarantee AI citations." Research analyzing 815,000 prompt-page pairs found that running the same ChatGPT prompt three times left only about 2.3% of citations consistent. You cannot guarantee an output that unstable. "We'll get you cited in AI answers" is the 2026 version of "we'll rank you #1."

The llms.txt problem

This one deserves its own section, because it's the most common paid deliverable in AI search packages and the easiest to verify.

What it is: a Markdown file at your domain root listing your important pages, proposed by Jeremy Howard in September 2024.

What Google says: it is not processed in any special way and is not a ranking signal — named explicitly in Google's May 2026 AI search guidance as a tactic site owners can skip.

What the crawl data says: adoption studies across hundreds of thousands of domains found the overwhelming majority of llms.txt files were never fetched at all, with no measurable citation effect.

Where it's legitimately useful: developer documentation and API references consumed by coding agents. Anthropic, Stripe, and Vercel all publish one — for that audience.

You will encounter vendor pages claiming Anthropic and Perplexity have "publicly confirmed support." That framing conflates a recommendation for agent-facing developer docs with consumption as a citation signal. They are not the same thing, and the difference is the entire product.

For an estate planning attorney in Mount Pleasant, an llms.txt file costs nothing to auto-generate and does nothing measurable. It should never be a line item.

Claims that are materially overstated

"SEO is dead — AI is where your customers are now." Measured LLM referral traffic runs under 2% of total referral traffic on average, with individual engines landing somewhere between 0.15% and 1.5%. The channel is growing fast and the traffic converts well. It is not a replacement for search, and any pitch that opens by declaring SEO dead is telling you what it needs you to believe rather than what the data shows.

There's a wrinkle worth knowing: the larger AI surface is inside Google itself. AI Overviews and AI Mode generate more AI-influenced traffic than all standalone chatbots combined — and they draw from the same index as organic search. Which means the "different discipline" framing collapses. Google's own documentation treats AEO and GEO as part of SEO.

"Our dashboard shows your AI visibility score." Give three visibility tools the same prompts and you'll get three different answers. Most run synthetic prompt sets the vendor curated, scored by a formula the vendor won't show you, measured against a baseline the vendor established. Point-in-time snapshots get presented as trend lines.

Prompt tracking isn't worthless — it just requires repeated runs and stated confidence intervals, which is exactly what a single blended "visibility score" hides. Ask to see the prompt set. Ask how many runs per prompt. If the answer is one, you're being sold luck.

"Allow the AI crawlers and you'll get cited." Necessary, not sufficient. Crawl access and citation are separate systems with separate user agents. We've reviewed server logs showing an engine crawling a site cleanly for months with zero citation fetches from the same vendor's answer bots. Unblocking is table stakes. It is not a growth strategy, and it is a two-hour fix, not a retainer.

"LLM traffic converts at 30–40%." The underlying studies are real. The interpretation is not. Someone who clicks through from an AI answer has already finished comparison shopping — the AI did the filtering. That's selection bias measured on very small denominators. It tells you the channel is high-intent. It does not tell you the channel is large.

The claim that's true — and the trap inside it

Reddit now accounts for roughly 40% of all LLM citations, ahead of Wikipedia and YouTube. Third-party mentions, consistent entity information across the web, and clean question-shaped content genuinely do influence whether you show up in AI answers.

That part is real. It's also the least novel part of the pitch — it's digital PR and content architecture wearing a new acronym.

The trap is what some vendors do with it. "Reddit placement" and "community seeding" services are astroturfing. For a law firm, that's a state bar advertising problem. For a fee-only RIA, that's an SEC Marketing Rule problem — testimonial and endorsement rules attach to compensated third-party statements regardless of where they appear. For a CPA firm, it's an AICPA independence and integrity question.

A marketing invoice does not convert a compliance violation into a marketing expense. If a vendor proposes it, ask them to put the compliance analysis in writing and route it to your bar counsel or CCO before anything ships.

What actually holds up

Strip out the theater and a short list survives:

  • Be indexable and crawlable. AI answers still draw predominantly from pages already ranking in the organic top twenty. Traditional search health remains the foundation, not a legacy concern.
  • Be a coherent entity. Consistent name, address, phone, credentials, and practice description everywhere you appear. This is what disambiguates you from the other three firms with your surname.
  • Write answer-first. A question as a heading, followed immediately by a direct answer in the first paragraph. Self-contained passages that make sense when lifted out of the page.
  • Earn real third-party mentions. Bar association directories, professional associations, legitimate press, industry publications.
  • Measure what's observable. Server logs for answer-engine fetches. GA4 referral segments for the engines that pass a referrer. Repeated prompt runs with the variance stated, not hidden.

None of that requires a new discipline. All of it is verifiable.

Three questions before you sign

  1. What result would prove this didn't work? If there's no answer, there's no accountability.
  2. What can you show me in my own server logs? Real crawler and fetch activity is observable on your own hardware. A vendor who can't point to it is asking you to trust their dashboard instead of your evidence.
  3. Which of these claims would you put in writing? Sales language and contract language often diverge sharply. The gap is informative.

How we approach this

SEO Doctor SC runs diagnostic audits only. We don't implement, we don't take the keyboard, and we don't sell the fix we just found. Every finding in our reports carries a confidence tier — Verified, Directional, Pending, or Unconfirmed — so you can see exactly which conclusions rest on hard evidence and which rest on inference.

If a vendor has handed you an AI search proposal and you want an independent read on what's real in it, that's precisely the kind of thing a flat-fee diagnostic is for.

Sources

Last reviewed: August 9, 2026. AI search behavior changes quickly; claims in this post are dated to the sources above.