AI SEO Agency: Vet One With Google's Own Checklist

Google publishes the questions you should ask before hiring anyone for AI search work. They sit on two pages in Search Central, they are specific, and they include this line: "If they have advice on optimizing for AI experiences (also known as 'AEO' 'GEO' services), is their advice aligned with Google Search's official guidance on optimizing for generative AI features?"

I read the ten pages currently ranking for "ai seo agency" in the United States. Four are agency service pages, three are agency roundups, one is a Reddit thread listing agencies. Between them they run roughly 16,000 words. Not one of them links to that question, and two of them sell things Google's own guidance lists as unnecessary.

That gap is the whole point of this article. You do not need an insider to evaluate an AI SEO pitch. You need one Google page, one browser tab of your own server logs, and about twenty minutes. No one can guarantee a #1 ranking on Google, and any agency that tells you otherwise has already failed the test.

What an AI SEO agency actually sells

Three different businesses use the same label, and the price difference between them is roughly tenfold.

The first is a conventional SEO agency that added an AI page to its service list. Its work is what SEO work has always been: technical audits, content briefs, digital PR for links. The AI part is a section in the deck. This is most of what ranks for the term, including the two highest organic results, both of which are /Services_Page templates on domains rated 80 and 65 by Ahrefs.

The second is a content factory. It sells volume: forty articles a month, generated and lightly edited, priced per piece. Whether that helps or hurts depends entirely on whether the pages are worth reading, a point I come back to below because Google's spam policy is written around exactly that distinction.

The third is a monitoring product wearing agency clothes. It tracks how often your brand appears in ChatGPT, Perplexity or Gemini answers and reports the number back to you. Useful, and completely separate from anything that changes the number.

The reason the distinction matters is that the three fail differently. An audit shop cannot fix thin content. A content factory cannot fix a site that assistants cannot crawl. A tracker fixes nothing at all and is often sold as if it did. Ask which of the three you are buying before you ask the price.

Google published the vetting questions, and the roundups do not use them

Search Central's page on hiring an SEO carries a list of interview questions. They are ordinary and they are effective, because a vendor who cannot answer them clearly is telling you something. Verbatim, the useful ones:

• "Can you show me examples of your previous work and share some success stories?"
• "Do you follow the Google Search Essentials (previously known as Webmaster Guidelines)?"
• "What kind of results do you expect to see, and in what timeframe? How do you measure your success?"
• "What's your experience in my industry?"
• "Will you share with me all the changes you make to my site, and provide detailed information about your recommendations and the reasoning behind them?"

Three more from the same page apply directly to anything labelled AI SEO. "Do they cite official Google documentation as supporting evidence for their recommendations?" "Is their advice aligned with Google Search's official guidance on optimizing for generative AI features?" "Do they use tools that are aligned with Google's guidance?"

The page is blunt about tooling too: "If your SEO uses a third-party tool, keep in mind that Google doesn't evaluate or endorse third-party SEO tools, and these tools don't have access to Google's internal ranking data. Be wary of tools that claim to be 'acceptable' or 'approved' by Google Search."

There is a matching warning inside the AI guidance itself: "Be wary of third-party tools that promise ranking success or claim to use 'internal' Google metrics. No third-party tool has access to our internal ranking or AI systems." That sentence invalidates a common demo, the dashboard that claims to show your "AI visibility score" as if it were a reading taken from Google.

None of this requires you to know SEO. It requires you to have read one page that the vendor probably assumes you have not.

The five things Google says you do not need

On 2026-07-10 Google published a guide to optimizing for generative AI search with a section titled "Mythbusting generative AI search: what you don't need to do". It is the cheapest vetting instrument available, because it lists five things by name. If a pitch is built on them, the pitch is built on sand.

Two clarifications, because each of these gets misquoted in both directions.

On llms.txt, Google is neutral rather than hostile: creating one "will neither harm nor help your site's visibility or rankings in Google Search, as Google Search ignores them." Other systems may read it. Google does not. So a proposal that treats the file as the deliverable is charging you for a file with no known consumer, which is the argument I made at length in our piece on LLM SEO.

On structured data, Google still recommends it: "it's a good idea to continue using it as part of your overall SEO strategy, as it helps with being eligible for rich results on Google Search." Markup earns you rich results. It does not earn you AI citations. Anyone selling schema as the path to being quoted by an assistant has the causality backwards, though the markup itself is worth having, as in the dentist schema we walked through line by line.

The same guide settles the naming argument that half these agencies are built on: "From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO."

Where the pitches and the documentation collide

I did not have to search for an example. The sixth organic result for this keyword is a 4,200-word agency roundup that ranks itself first. Two of its recommendations sit directly against the guide above.

It tells readers to "prioritize technical SEO and structured-data depth, and ask how the firm makes content legible to large language models." Google's mythbusting list says structured data is not required for generative AI search and that you do not need to write in a specific way for AI systems, because "AI systems can understand synonyms and general meanings of what someone is seeking."

It also states, without a source: "Plan on three to six months for early movement in AI Overviews and assistant citations, and six to twelve months for results that show up in referral traffic and revenue." Google's SEO starter guide says: "Some changes might take effect in a few hours, others could take several months. In general, you likely want to wait a few weeks to assess whether your work had beneficial effects in Google Search results."

Neither of those makes the agency dishonest. It makes them unsourced, which is the specific thing Google's vetting question asks you to check. Ask for the citation. Watch what happens.

What actually decides whether an assistant can read you

Crawler access, and it is not one setting. Each vendor runs several agents with different jobs, and the robots.txt contract differs per agent and per vendor. This is the layer no roundup in the top ten mentions, and it is the layer where a site silently disappears.

OpenAI documents four agents. OAI-SearchBot surfaces pages in ChatGPT search, and OpenAI recommends allowing it. GPTBot crawls content for model training. ChatGPT-User fetches a page when a user asks ChatGPT about it, and here is the sentence that matters: "Because these actions are initiated by a user, robots.txt rules may not apply." OAI-AdsBot only visits pages submitted as ads.

Anthropic documents three: ClaudeBot for training, Claude-User when someone asks Claude about a page, and Claude-SearchBot for search quality. Anthropic states that its bots "respect 'do not crawl' signals by honoring industry standard directives in robots.txt", and that includes the user-initiated agent.

So the two vendors take opposite positions on the equivalent agent. One blanket rule cannot express what you want across both. A file that says what you actually mean looks like this:

# Allow retrieval for search and answers, refuse model training.
User-agent: OAI-SearchBot
Allow: /

User-agent: Claude-SearchBot
Allow: /

User-agent: Claude-User
Allow: /

User-agent: GPTBot
Disallow: /

User-agent: ClaudeBot
Disallow: /

User-agent: Google-Extended
Disallow: /

User-agent: *
Allow: /

Sitemap: https://example.com/sitemap.xml

Note what that file does not do: it does not block Googlebot, and blocking Googlebot is how you leave Google's AI features entirely. The guide is explicit that a page must be indexed and "eligible to be shown in Google Search with a snippet" before it can appear in a generative AI feature at all. A stray nosnippet from the 2024 panic about AI training is enough to remove a page from AI Overviews while leaving its blue link untouched. If your site runs behind a CDN with an "AI crawler" toggle, check what that toggle covers before an agency does it for you. We covered the Cloudflare version of that trap in the LLM SEO article.

Run the checks yourself before the sales call

Every claim above is verifiable from a terminal in a few minutes. Do it before you take the meeting, because it changes what you ask.

First, see what you are currently telling each agent:

# robots-check.sh
DOMAIN="example.com"
curl -s "https://$DOMAIN/robots.txt" | grep -iE \
  'OAI-SearchBot|GPTBot|ChatGPT-User|OAI-AdsBot|ClaudeBot|Claude-User|Claude-SearchBot|Google-Extended|PerplexityBot|Googlebot'

Second, check whether the page has any text in the HTML before JavaScript runs. Assistant retrieval fetches your page like a simple client, so a page whose body arrives empty is a page that gets summarised as nothing:

curl -s "https://example.com/your-page" \
  | sed -e 's/<script[^>]*>.*<\/script>//g' -e 's/<[^>]*>/ /g' \
  | tr -s ' \n' ' ' | wc -w

If that number is close to zero while the page looks full in a browser, you have a rendering problem, not a content problem. Our Next.js SEO walkthrough covers what the App Router does and does not put in the initial HTML.

Third, find out whether these crawlers are visiting at all. Every vendor publishes its IP ranges, and all five files were live when this article was written:

for f in searchbot gptbot chatgpt-user adsbot; do
  echo "== openai/$f"
  curl -s "https://openai.com/$f.json" | head -c 200; echo
done
curl -s "https://claude.com/crawling/bots.json" | head -c 200

Fourth, grep your access log for the agents themselves. This is the number an agency should be asking you for in the first meeting, and almost none do:

# ai-crawler-hits.sh
LOG="/var/log/nginx/access.log"
for ua in OAI-SearchBot GPTBot ChatGPT-User ClaudeBot Claude-User Claude-SearchBot PerplexityBot Googlebot; do
  printf "%-20s %s\n" "$ua" "$(grep -c "$ua" "$LOG")"
done

A vendor who wants to sell you AI visibility and has never asked to see that output is selling a strategy without a baseline.

What it costs, with numbers you can check

Published figures, all from the vendors' own pages or a named survey, none invented.

Two caveats on that last row, and they matter. The Ahrefs survey is dated "Updated: August 15, 2024" and it covers SEO generally rather than AI search work specifically. It is a calibration point, not a quote.

The spread between $1,000 and $50,000 is not a spread in quality. It is a spread in what is inside the retainer: strategy hours, technical audit, content production, link acquisition, reporting. We took a retainer apart line by line in the SaaS SEO agency piece, and the arithmetic there transfers directly. Ask for the split before you compare two quotes, because two identical monthly numbers can contain entirely different work.

One more thing worth pricing separately: production. If the proposal is mostly articles, divide the retainer by the number of articles and compare that per-article number against everything else on the invoice. It is usually the largest line and the easiest one to move.

How to measure whether any of it worked

Set the measurement up before the engagement starts, so the baseline is not the agency's own screenshot.

Search Console now has a Generative AI performance report. Its definition of an impression is precise: "Impressions are how many times links to your site were shown to a user in a generative AI feature on Google Search." It covers AI Overviews and AI Mode, broken down by page, country, date and device.

What it does not have is equally precise, and it kills a lot of promised dashboards: no clicks, no click-through rate, no position, no query data. Anyone quoting you an AI Overview click count for your site is not reading it from Google. Three documented reasons the report may be missing: the rollout has not reached your property, the site has not received enough impressions, or the site is excluded from Search generative AI features.

For assistants outside Google, the honest method is a frozen panel of prompts. Write twenty questions a buyer would actually type, run them on a fixed schedule against each assistant, and record whether your domain is cited and which domains are cited instead. It is crude, it is repeatable, and it is the only method that survives a vendor change. We described the mechanics in how to rank on ChatGPT.

Then add the boring layer that most AI SEO conversations skip: organic sessions, assisted conversions, and referral traffic from chatgpt.com, perplexity.ai and claude.ai in your analytics. Referral traffic from assistants is small for most sites and it is real, which makes it a better trend line than any score out of 100.

The claims that should end the call

Six of them, each tied to something published rather than to taste.

• A guaranteed position or a guaranteed timeline. Google: "No one can guarantee a #1 ranking on Google." And on audits: "If they guarantee you that their changes will give you first place in search results, find someone else."
• "Internal" or "approved" Google metrics. Google: "No third-party tool has access to our internal ranking or AI systems", and be wary of tools claiming to be "acceptable" or "approved" by Google Search.
• llms.txt sold as the deliverable. Fine to have, ignored by Google Search, and not a project.
• Mention farming. Paid placements engineered to make assistants repeat your name run into the site reputation abuse policy, which targets third-party content published on a host site "mainly because of that host's already-established ranking signals".
• Volume with no editorial standard. Scaled content abuse "is when many pages are generated for the primary purpose of manipulating search rankings and not helping users", with generative AI tools named as one method. The policy is about purpose, not about tooling, which is the distinction that decides whether AI-assisted publishing is fine or fatal.
• No baseline. If nobody asked for your Search Console access, your crawler logs, or your current AI Overview impressions before quoting, the proposal is a template.

A vendor who fails one of these is not necessarily a fraud. A vendor who fails three has not read the documentation for the thing they are selling.

Agency, tool, or your own hands

The honest answer depends on which resource you are short of, and it is worth saying out loud that the three are not competitors so much as different line items.

Google's own position on that last row is worth quoting because it is the opposite of what most of this SERP sells: "If you run a small local business, you can probably do much of the work yourself. We recommend starting with the SEO starter guide." We wrote the service business version and the independent professional version of that argument, and both hold.

The decision that trips people up is choosing between an agency and a tool when the actual bottleneck is a third thing. If your pages are not indexed, more pages do not help. If your positioning is vague, better crawling does not help. Diagnose the bottleneck first, then buy for it.

Where we fit, and where we do not

syted is the production line item, not the agency. It researches a keyword, checks the live search results against what you sell, writes one article a day and publishes it on your own domain at yourdomain.com/blog. That is $99 per month for thirty articles, one plan, cancel in one click.

What it does not do: technical audits, link building, ad campaigns, or reporting on rankings. If your bottleneck is any of those, an agency is the right purchase and this is not. The SEO versus Google Ads breakdown walks through the case where paid is the better answer entirely.

The reason we can describe the audience check honestly is that it runs on this blog too. Before this article existed, the keyword went through the same gate every customer keyword goes through: open the real search results, read who ranks and what those pages are actually known for, and skip the keyword when the audience does not match. That check is also why our SaaS SEO article exists and why several higher-volume keywords do not.

Nobody can promise you a position, and results vary with your site, your category and your competition. What you can insist on is that every recommendation you pay for cites something you can open and read. That is Google's own vetting question, and it costs nothing to ask.

FAQ

What is an AI SEO agency?

A firm selling work aimed at visibility in AI-generated search results: AI Overviews and AI Mode on Google, plus answers from ChatGPT, Claude, Perplexity and Gemini. In practice the label covers three different businesses, a conventional SEO agency with an AI service page, a content production shop, and a monitoring tool. Google's position on the underlying work is that "optimizing for generative AI search is optimizing for the search experience, and thus still SEO."

Can AI agents do SEO?

Parts of it, reliably. Production scales well: research, drafting, publishing, and the repetitive technical checks in this article. Judgement does not, at least not unattended. Deciding which keyword is worth a page, whether a page belongs on your site at all, and whether a claim is true still needs a person who knows the business. Google's spam policy draws the line at purpose rather than tooling: pages "generated for the primary purpose of manipulating search rankings and not helping users" are the violation, no matter how they were made.

How much does AI SEO cost?

Published ranges from vendors' own pages in 2026 run roughly $1,000 to $5,000 per month for small and local businesses, $5,000 to $15,000 for mid-market, and $15,000 and up for enterprise, with one roundup citing $50,000 and above at the top end. For general SEO, an Ahrefs survey of 439 providers dated August 2024 put agency averages at $3,209 per month and freelancers at $1,348. Compare the split inside the retainer rather than the headline number.

What is the best AI SEO agency?

There is no defensible single answer, and be careful with the roundups that offer one: several of the pages ranking for this question place themselves at number one on their own list. A more useful question is which of the three business models above matches your bottleneck, and whether the vendor can cite official documentation for its recommendations, which is the check Google itself publishes.

Do I need a separate AI SEO agency, or can my current one do it?

Ask your current agency the AI-specific vetting question from Search Central: is their advice aligned with Google's official guidance on optimizing for generative AI features? If they can name that guide and say which of its five mythbusted items they avoid, you likely do not need a second vendor. If the answer is a deck about llms.txt and content chunking, you do not need a second vendor either, you need a different one.

Does an llms.txt file get me cited by AI assistants?

Not by Google Search, which states plainly that it does not use such files and that creating one "will neither harm nor help your site's visibility or rankings in Google Search." No major assistant vendor has published a statement that its production retrieval reads third-party llms.txt files. Publishing one is harmless and cheap. Paying an agency for one as a deliverable is neither.