AI SEO Agencies: What They Do and How to Vet One

Most AI SEO agencies sell some mix of six services, whatever they call the work (GEO, AEO, AI SEO or LLM optimization):

• Visibility tracking: running a set of prompts through ChatGPT, Perplexity, Gemini and others, then reporting how often your brand appears.
• Technical access audits: checking whether AI crawlers can fetch and read your pages.
• Content rewrites "for AI": breaking pages into chunks, adding statistics or quotations, or writing in a "quotable" style.
• Markup and files: schema, FAQ schema and llms.txt.
• Off-site work: getting your brand named on the sites engines already cite.
• Accuracy work: finding where AI answers state the wrong price, product or fact about you, and fixing the source.

The labels differ. The menus look alike. What separates agencies is how much of the menu has evidence behind it.

One finding sorts most of it. A critical survey of 45 GEO studies (arXiv 2607.14035, July 2026) concluded that content already retrieved can change its own citation, but no reviewed technique shows a stable, cross-platform causal effect on whether a page gets discovered. Most of the market sells the second thing. The evidence supports only the first.

An AI answer about your category passes through four stages. Each has a different owner.

1. Parametric: what the model learned in training. It is frozen at a knowledge cutoff. Nobody changes it this quarter.
2. Access: whether a crawler may fetch and store your page. You and your CDN control this.
3. Retrieval: whether your page enters the model's context for a given question. The engine controls this, and it is per query. As of October 2026, ChatGPT decides when to search, and Google AI Overviews often do not appear at all.
4. Generation: which retrieved pages get used and cited. The model controls this.

Almost every published GEO technique was tested at stage 4, on pages already placed in the model's context. The best-designed of those studies, C-SEO Bench (NeurIPS 2025), found most rewriting methods "largely ineffective" and often harmful to the page's ranking in the context.

So an agency selling rewrites is selling a stage 4 result as a stage 2 outcome. The strong evidence sits at stage 2, access. It works by removing blocks, not by adding signals.

Here is how common agency services line up against primary sources.

One caveat on schema. Every page in the Ahrefs study already had 100 or more AI Overview citations, so whether schema helps a page enter the set is untested. Google also says existing markup does no harm. Keep what you have. Don't pay for it as an AI lever.

Treat any "up to 40% more visibility" claim the same way. That figure comes from a lab engine the GEO paper's authors built, with sources already loaded into context.

Be clear about the split before you hire anyone.

In your control (or your agency's)

• Whether key content is in the HTML before JavaScript runs.
• Per-bot robots.txt rules. Blocking GPTBot stops training use. OAI-SearchBot controls ChatGPT search.
• Firewall, CDN and bot-management rules that may block a crawler robots.txt allows.
• noindex, nosnippet, max-snippet and data-nosnippet on your pages.
• Google's generative AI inclusion setting in Search Console, which is on by default and can be inherited from a parent property.
• Whether your own pages state prices, products and facts consistently.

A typical robots.txt that keeps ChatGPT search while opting out of training looks like this:

User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Disallow: /

Out of everyone's control

• What the model learned before its cutoff.
• Whether the engine searches for a given question, and how it rewrites the query.
• Where your page lands in the model's context, and which sources it cites.
• The pull of brand size. Large brands dominate citations in every study, and no study controls for it.

An agency can fix the first list. It can only influence the second list indirectly, and it can't promise anything about it.

Whether you hire or do it in-house, work in this order. Each step removes a reason the next one could fail.

1. Baseline each platform separately. Ask the buyer questions your customers ask, one engine at a time. The free checker on this site shows whether engines name you for up to three buyer questions, as a one-off snapshot.
2. Check access. Read robots.txt bot by bot. Then request your pages as each AI crawler, because a firewall can return 403 to a bot robots.txt allows. A free AI crawler checker does this. Since 15 September 2026, Cloudflare blocks Training and Agent crawlers by default on ad-displaying pages for new domains and free-tier customers, so a site can lose access without anyone editing it.
3. Check rendering. Compare the raw HTML with the rendered page. Prices, product names and core facts must appear before JavaScript runs.
4. Check eligibility. Look for inherited noindex or nosnippet tags. Check Google Search Console and Bing Webmaster Tools separately, since one healthy index says nothing about the other.
5. Fix wrong facts. List what AI answers get wrong. Remove contradictions on your own site, and state each fact in one plain sentence.
6. Work off-site. Where engines cite other sites for your category, aim to be named on those sites.
7. Re-measure. Use the same questions, logged out, per platform.

Small teams can cover steps 1 to 4 in the one-afternoon AI visibility check before deciding whether they need outside help.

These questions separate agencies that fix things from agencies that sell claims. A fuller list sits in choosing an AI SEO agency.

• "Do you test our firewall as well as robots.txt?" A good answer names the crawlers and describes requesting pages as each one.
• "What's the difference between GPTBot and OAI-SearchBot?" A wrong answer is "allow GPTBot so ChatGPT can find you." GPTBot governs training only.
• "How do you check what our pages show before JavaScript runs?" Look for a raw-HTML versus rendered comparison.
• "Which recommendations have primary-source evidence, and which are your judgment?" Honest agencies can split the two.
• "Do you report per platform, and do you note whether the engine searched?" A single number hides which engine is failing. See one score versus per-platform scores.
• "What do you guarantee?" The right answer is delivered fixes. Anyone promising citations is promising something no one controls, as covered in whether to trust citation guarantees.
• "Can we see the prompts and raw answers behind your case studies?" Without them, you can't separate engine behavior from prompt choice. Read how to judge GEO case studies.
• "Do you report an AI rank or position?" As of October 2026, Google's AI report shows impressions only, so no position figure can be checked.

• Shortlisting from a ranked list. Many lists are written by the agencies on them. See whether to trust top agency lists.
• Paying for a schema and llms.txt bundle. It is the most common package and the least supported.
• Accepting one blended score. Engines behave differently. One may cite you often while others ignore you.
• Mass-publishing AI content. In OtterlyAI's 2026 experiment, Google removed two new sites carrying about 1,000 AI-generated posts each, with no manual action reported in Search Console. Google's spam policy calls this scaled content abuse.
• Assuming your SEO agency already covers it. Crawler access and AI accuracy are often outside a standard retainer. Ask, as explained in whether your SEO agency should handle AI search.
• Tolerating manipulation. Hidden text, prompt injection and cloaking for AI crawlers exist, and research shows some engines can be swayed by them. They are illegitimate and expose your brand to penalties and reputational damage. Walk away from any agency that suggests them.

Is an AI SEO agency different from a regular SEO agency? Partly. Google says there are no special requirements to appear in AI Overviews or AI Mode beyond being indexed and snippet-eligible, so good technical SEO already covers Google. The difference lies in non-Google engines. Their crawlers mostly don't run JavaScript, each has its own robots.txt token, and firewalls can block them. An agency that tests those gates and fixes wrong facts in AI answers adds something a standard retainer may miss.

Should an agency add schema markup to help us appear in AI answers? Not as an AI lever. Google says structured data isn't required for generative AI search, and Ahrefs' matched test found AI Overview citations fell 4.6% on pages after schema was added. FAQ rich results were deprecated in May 2026. Existing markup does no harm, so there's no need to remove it. Just don't pay for schema work sold as a way to get cited.

Can an agency guarantee AI citations or a visibility score? No. Engines decide per question whether to search, rewrite your query, and choose which retrieved sources to cite. None of that is in an agency's control. What an agency can guarantee is work: crawler access tested and fixed, key content moved into the HTML, accidental noindex or nosnippet tags removed, and wrong facts corrected on your site. Judge them on those deliverables.

How should an agency report AI visibility results? Per platform, with the questions listed and the date run. Reports should say whether each answer used live search, since answers with and without retrieval behave differently. They should show wrong facts found and fixed, and access checks passed. Be wary of an AI rank or position for Google. As of October 2026, Google exposes only impressions for its AI features, so no position number can be verified.

Should we block GPTBot to protect our content? That's a business choice, and it won't remove you from ChatGPT search. OpenAI documents GPTBot as a training crawler and OAI-SearchBot as the search crawler, and says each setting is independent. You can disallow GPTBot and allow OAI-SearchBot. Then confirm your CDN or firewall lets OAI-SearchBot through too, since OpenAI says eligibility needs both.