Choosing the Best AI SEO Agency: Questions to Ask

Most AI SEO pitches sound alike. The differences show up when you ask how the agency measures results and what it believes moves them.

The evidence base is thin, and good agencies know it. A critical survey of 45 GEO studies (arXiv, July 2026) concluded that "already-retrieved content can causally alter its citation or use, but no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior." In plain terms, tweaks can change how a page is used once an engine has already picked it up. Nothing has been shown to reliably get a page picked up in the first place.

That is why a ranked list of agencies can't make this choice for you. We cover the limits of those lists in whether you can trust top AI SEO agency lists. Use the questions below instead. They test whether an agency's claims match what primary sources say. For a broader view of what these firms do, see what AI SEO agencies do and how to vet one.

You need this model to judge any agency's answers. Every AI answer passes through up to four stages:

• Parametric: what the model knows with no web access, frozen at a training cutoff. Nobody changes this in a given quarter.
• Access: whether a crawler may fetch and store your page at all. You and your CDN control this.
• Retrieval: whether your page enters the model's context for a specific question. The engine controls this, partly through its own query rewriting.
• Generation: which retrieved pages the model uses and cites. The model controls this.

Retrieval is not constant. As of October 2026, ChatGPT decides per question whether to search, Claude.ai search is a user toggle, and Google AI Overviews often don't appear at all. So any single visibility number blends answers given with and without live search. An agency that doesn't say how it handles that mix is reporting an average of two different things.

Citation is also not the same as influence. Engines retrieve more pages than they cite, and research on how models cite (arXiv, 2026) shows citation output can be changed without changing the sources. Treat citation counts as a useful but noisy signal.

Sort every proposed task into one of these two lists before you sign.

In your control, and so in theirs

• Per-bot crawler permission. OpenAI's crawler page says each bot setting "is independent of the others". GPTBot governs training. OAI-SearchBot governs ChatGPT search.
• Network reachability. A firewall or CDN can block an AI crawler that robots.txt allows. From 15 September 2026, Cloudflare blocks training and agent crawlers by default on ad-displaying pages for new domains and existing free-tier customers.
• Rendering. Vercel found ChatGPT and Claude crawlers fetch JavaScript files but don't execute them, so content that only appears after JavaScript runs may be invisible to them.
• Index and snippet eligibility. Google says a page must be indexed and eligible for a snippet to appear as a supporting link in AI Overviews or AI Mode, with "no additional technical requirements."
• Accidental suppression. noindex, nosnippet, max-snippet:0 and data-nosnippet keep content out of AI answers.
• Stated facts. A specific price, number or one-line definition on a fetchable page gives an engine a sentence it can quote.

Not in anyone's control

• What the model learned in training.
• Which queries the engine actually runs.
• Where your page lands among the retrieved results.
• Whether the engine searches at all for a given question.

Brand prominence sits between the two. Being named on the sites engines already cite is associated with appearing in answers. It is not a guaranteed effect, and an honest agency will call it an association.

Send these in writing and compare the replies side by side.

1. Do you record whether each answer used live search? A good answer explains how they control or disclose retrieval state. A weak answer is a single blended score with no method.
2. Do you report each engine separately? Google says AI Overviews and AI Mode "may use different models and techniques". Blending engines hides where you actually stand.
3. Where do your test prompts come from? Engines rewrite questions into their own search queries. As of October 2026, Bing Webmaster Tools is the only first-party source that shows grounding queries. A good agency treats its prompt set as a sample, built with a method like structured prompt research.
4. Do you report a rank or position inside AI answers? Google's generative AI report in Search Console exposes impressions only. There is no position data to check such a number against.
5. How do you separate a mention, a citation and a retrieval? These are different events, and the gap between them matters.
6. Do you test logged out with memory off? Personalization makes logged-in results hard to reproduce.
7. Which of your numbers are measurements, and which are your model? The best agencies answer this without hesitation.

Once you know what they measure, agree on metrics you can defend to your board. Our guide to AI visibility KPIs covers which ones hold up.

Measurement tells you what they see. These questions tell you what they'll do about it.

1. What do you check first? Look for access work: robots.txt per bot, firewall and CDN reachability, rendering, index status in both Google and Bing. If the first deliverable is content or markup, ask why.
2. Which bot controls ChatGPT search? The right answer is OAI-SearchBot. An agency that says "allow GPTBot so ChatGPT can find you" has the mechanism wrong.
3. How do you check what AI says about us is accurate? AI platforms get specific facts wrong about specific brands. Ask how they find wrong prices, outdated products or mixed-up competitors, and how they trace each error to a fix.
4. Do you publish content at volume? OtterlyAI tested roughly 1,000 AI-generated posts on each of two new domains in 2026. Google deindexed both without a manual action, one after about 14 days. Google's spam policy calls this scaled content abuse.
5. Would you ever use hidden text or AI-only content? The only acceptable answer is no. Prompt injection and cloaking for AI crawlers exist, and they put your brand at risk.
6. What do you think of schema, llms.txt and E-E-A-T for AI? Compare their answer with the table below.

These claims show up in many pitches. Here is what primary sources say.

Don't overcorrect. Google says existing markup does no harm, so an agency that tells you to rip out schema is also wrong. The honest position is that it isn't a lever for AI answers.

1. Baseline yourself first. Before any call, see where you stand. The free checker on this site shows whether AI engines name you for up to three buyer questions. It is a snapshot, which is enough to judge an agency's first report.
2. Check crawler access. The AI Crawler Checker requests a page as each AI crawler and shows which ones your firewall turns away. If an agency's audit misses a block you already found, that tells you something.
3. Send the questions in writing. Ask every shortlisted agency the same list, so the replies are comparable.
4. Ask for a sample report. Check that it separates engines, discloses retrieval state and avoids rank language.
5. Read the contract for promises. Any guarantee of a citation, position or score belongs in the red-flag pile. See why citation guarantees don't hold.
6. Set a review point. Agree which access fixes ship first and how you'll confirm each one shipped.

• Buying the biggest promise. Agencies that guarantee outcomes are promising something nobody controls.
• Accepting one blended score. It hides which engine you're missing from and why.
• Skipping access work. A blocked crawler or a JavaScript-only pricing page can cancel every other task.
• Assuming your SEO agency already covers this. Some do, many don't. Check with the questions in whether your SEO agency should handle AI search.
• Paying for volume content. Otterly's test shows how quietly Google can remove a site built that way.
• Hiring before checking in-house capacity. Much of the access work is a one-time technical fix. Weigh that against doing it in-house.

How much should an AI SEO agency cost? No published benchmark sets a fair price, so judge value by scope. Ask what share of the work is access fixes, accuracy checks and measurement, and what share is content or markup the evidence doesn't support. A smaller retainer focused on access and accurate facts may do more than a large one built around schema and volume content. Ask for deliverables tied to specific fixes, not to citation targets.

Can an agency get my brand into ChatGPT answers? It can remove blocks, but it can't guarantee inclusion. An agency can make sure OAI-SearchBot is allowed, your firewall lets OpenAI's crawler through, and your key facts are in the HTML before JavaScript runs. Those steps make you eligible. Whether ChatGPT searches, retrieves your page and cites it stays with OpenAI. Any agency that promises the outcome is overstating what it controls.

What should an AI visibility report from an agency include? Results per engine, not one blended number. A note on whether answers used live search. The prompt list and how it was built. Mentions, citations and accuracy errors kept separate. A list of access issues found and fixed, with dates. It should not include an AI rank or position, because no engine publishes data that could validate one.

Is an agency that recommends schema markup wrong? Not entirely, but check the reason they give. Schema can still serve classic search features, and Google says existing markup does no harm. The problem is selling it as a way to get cited in AI answers. Google says structured data isn't required for generative AI search, and Ahrefs' matched test found AI Overview citations fell 4.6% after pages added it.

How long before I see results from an agency? Access fixes show up quickly in technical checks: a crawler that got a 403 now gets a 200, or pricing now appears in raw HTML. Whether answers change afterward depends on when engines recrawl and whether they retrieve your page for a given question. No source measures a reliable timeline. Treat any agency that quotes one with caution, and track per-engine results over several months.