How to Choose an AI SEO Agency

A brand appearing in an AI answer is an observation, not an explanation. Treating those two things as interchangeable is where expensive mistakes begin.

A useful AI-search process should be able to move through a chain like this:

Question → Observation → Evidence → Interpretation → Action → Recheck

Each part has its own job. The question establishes what was tested. The observation records what actually happened. The evidence preserves enough of the result for someone else to inspect it. The interpretation explains what the agency thinks the result means. The action is the recommendation that follows, and the recheck is how you find out whether the decision changed anything.

Search work gets expensive when interpretation starts masquerading as fact.

Consider a simple example. An agency tests a commercially important question and your company does not appear in the AI answer. A competitor does, and the answer cites a guide from that competitor’s website.

The observation is straightforward: your company was absent, the competitor appeared, and one of its pages was cited. The preserved evidence might include the exact question, platform, date, answer, cited source, URL and result type.

The difficult part starts after that.

One explanation is that the competitor’s page is stronger. It could also have a more established brand, better traditional rankings, more third-party references, clearer topical coverage, fresher information or simply a page that was more retrievable for that particular answer at that particular moment.

The weak response is to jump straight from “they have a guide” to “we need a guide.”

A better response asks what the evidence actually supports. If your site has no page that credibly answers the subject, creating or improving content may be justified. If you already have a strong page, publishing another could split ownership and make the problem worse. If the existing page is difficult to crawl or index, more copy is probably not the first move. If the competitor’s advantage appears to come mainly from third-party authority, another page on your own domain may do very little.

There are also situations where I would refuse to approve a large change because the diagnosis has not earned it yet. A technically plausible recommendation is not enough if the evidence is thin, the downside is meaningful and the proposed work would be difficult to reverse. In that situation, narrowing the intervention, preserving the competing explanations and collecting another measurement is a better decision than manufacturing certainty because a deliverable has already been scoped.

The explanation needs to separate what was observed from what was inferred, account for plausible alternatives and show why the recommended action still deserves the client’s money.

Our deeper guide to how AI search citations work gets into the mechanics behind mentions, citations and source selection without pretending that a citation automatically tells you why it happened.

Forty screenshots are not necessary to decide whether someone understands AI-search measurement. One finding, walked through properly from beginning to end, tells me much more.

The useful version is the working version, not only the polished case study after the result turned out well. A useful walkthrough shows what was tested, which surface produced the result, what actually appeared, which source or page was involved and what evidence was preserved.

Then comes the part that matters more: judgment.

Why does the provider think the result happened? How certain is that explanation? What else could plausibly explain it? What is being proposed because of the finding, and what evidence would make the team stop, change direction or leave the site alone?

I would put these questions to one real finding on a vendor call, before anyone opens the sales deck.

This is a demonstration standard, not a checklist for its own sake. If someone cannot take one real observation and carry it cleanly through interpretation, decision and recheck, I do not expect the reasoning to improve when the dataset becomes larger or the dashboard becomes more sophisticated.

Certainty everywhere would make me more skeptical, not less, because search rarely gives us that luxury.

Google makes a similar distinction in its guidance for evaluating outside SEO advice. It says third-party SEO tools do not have access to Google’s internal ranking data and cannot guarantee performance, and it recommends checking outside recommendations against official Search guidance rather than treating proprietary tools as privileged access to Google’s systems. (Google Search documentation)

Tools are useful. They are not oracles, and the interpretation still has to earn its way into the recommendation.

A visibility score can compress three different layers into one number: measurement, aggregation and interpretation. If I cannot separate those layers again, the score tells me less than the dashboard suggests.

Measurement is the underlying observation: whether the brand appeared, whether it was mentioned, cited or recommended, which page showed up, on what platform, for which question and on what date.

Aggregation compresses many of those observations into something easier to compare. That may be a score, percentage, share of voice, visibility index or another summary.

Interpretation is the conclusion someone draws from the pattern.

Those layers are related, but they are not interchangeable. If an agency tells you your “AI visibility score” is 42, the useful question is what produced the 42. The number only becomes meaningful once it can be traced back to things that actually happened.

A credible report should let you get underneath the aggregate and inspect the inputs:

• the question or prompt;
• the platform;
• the date;
• whether the brand appeared;
• whether it was mentioned, cited or recommended;
• the URL or source involved;
• the position or result type where relevant;
• what changed between measurement periods.

The score can help reveal a pattern. It should not make the underlying evidence disappear.

This is why we measure AI visibility before you optimize. A baseline does not need to be exciting. Its value is that it records what existed before the work began, before there is a success story to protect and before anyone knows which direction the numbers are going to move.

Our own Day 0 baseline follows the same principle. Recording the starting state before there is anything flattering to report gives later claims something concrete to be compared against.

If you want to go deeper into the measurement layer itself, Measuring AI Search Visibility explains those distinctions in more detail.

I keep the surfaces separate when I evaluate an AI SEO agency. Google AI Overviews, AI Mode, ChatGPT search, Perplexity and other answer systems do not necessarily retrieve, synthesize or cite information in the same way. Turning them into one number too early can hide the result that actually needs explaining.

Even Google’s own AI search experiences can use different models and techniques and return different responses and links. Google also states that standard SEO foundations remain relevant and that there are no special additional requirements for inclusion in AI Overviews or AI Mode. (Google Search documentation)

ChatGPT search exposes sources and citations, while OpenAI explicitly warns that search results and citations can be incomplete, outdated or incorrect. (OpenAI Help Center)

Perplexity similarly exposes sources and tells users to validate information against the linked material behind an answer. (Perplexity Help Center)

The buyer implication matters more than the platform trivia. If an agency combines all of those surfaces into one score, the report needs to make clear exactly what information disappears in the aggregation.

A blended metric can be useful for trend monitoring. It becomes misleading when it creates the impression that a brand has one objective “AI ranking” across every system.

We think about modern discovery as one search ecosystem, but that does not mean every surface behaves identically. They can belong to the same customer journey and still require separate observation because the weakness, and therefore the recommendation, may differ by surface.

“We don’t know” should not end the analysis. It should change the scale and risk of the next decision.

When evidence is weak, the claim should become narrower, competing explanations should remain visible and the next action should match the confidence actually available. That usually means choosing something cheaper, more reversible or useful under several plausible explanations, then measuring again.

For example, if we see a visibility gap but cannot isolate the cause, I may still approve an improvement to a weak page if that page is clearly underserving users and traditional search too. The action makes sense even if the AI-search hypothesis later turns out to be wrong.

Rebuilding a large section of a site because one AI response looked unusual is a very different proposition.

There are times when a technical correction is justified, times when content deserves improvement anyway, and times when more measurement is the right next step. Leaving the site alone can also be the correct decision.

I have been around search long enough to distrust strategies where every finding conveniently produces another billable deliverable. Useful analysis should occasionally save the client from work that does not deserve to happen.

That is one reason human judgment still has to run the process. Automation can collect an enormous number of observations. It cannot decide which observations deserve action, how much confidence they deserve or how much money should be put behind them.

One of the easiest ways to turn AI-search work into busywork is to treat every missing citation as a content gap. Sometimes the site genuinely lacks the information. Other times the information already exists and the real problem is somewhere else.

Creating another page can split topic ownership, compete with an existing page, introduce semantic redundancy and leave search systems with several weak candidates instead of one clearly authoritative page. It can also create the appearance of progress because something shipped even though the visibility problem itself remains unsolved.

Producing content merely because a measurement found something uncomfortable does not interest me. The question is whether another page improves the site’s ability to answer the reader’s need.

If it does, build it. If an existing page should own the subject, strengthen that page instead. If the problem is technical or external, fix the actual problem rather than producing content because content happens to be the easiest deliverable to sell.

Google’s current guidance for generative AI search points in the same direction. It emphasizes useful, unique, people-first content and warns against producing large numbers of pages for query variations simply to manipulate visibility. (Google Search documentation)

Shipping more pages is not a useful success metric by itself. The search system should become clearer and more useful because of the work.

Agencies offer schema, rendering and crawler access as proof of “AI readiness” all the time, so I check how far that evidence goes. Technical conversations about AI search often move too quickly from “the crawler can access us” to “we are optimized for AI.” A more useful way to evaluate the claim is to separate eligibility, visibility and influence.

Eligibility is about whether the search or AI system can access, index, retrieve or otherwise use the information. Crawler access, robots directives, rendering, indexing, textual availability and technical discoverability live here.

Visibility starts later. It asks whether the brand, page or source actually appears for the questions that matter, through mentions, citations, supporting links, recommendations or other observable search appearances.

Influence goes further again. Did that visibility contribute to something meaningful? Did it help a person discover the brand, reach the site, consider the business, convert or create measurable demand?

The three layers are connected, but one does not prove the next. A crawler reaching a page tells you the page is accessible to that crawler. Seeing the brand or page actually appear answers a different question. Traffic or conversion movement gets closer to influence, although attribution can still remain uncertain.

Keeping those layers separate prevents technical work from being oversold. Crawler access, JavaScript rendering, structured data and machine readability can matter a great deal when there is a real barrier, but they should not be presented as shortcuts around relevance, authority or usefulness.

Our guide to AI crawlers and discovery goes further into that eligibility layer.

Google similarly says there are no special schema.org requirements or new machine-readable AI files needed for inclusion in its AI search features. Pages still need to satisfy ordinary Search eligibility and indexing requirements. (Google Search documentation)

If terminology such as AEO, GEO and AIO is making the subject harder rather than clearer, the AI Search Field Guide and our explanation of AEO, GEO and AIO cover those definitions without turning this article into another acronym glossary.

Look closely at the recommendation itself. “Improve authority,” “create more content,” “add schema” and “build more citations” are activities. None of them explains what actually went wrong.

If the finding is important enough to spend money on, the recommendation should be traceable back to something concrete.

Suppose an audit finds poor AI visibility. The problem might involve content, crawlability, page ownership, weak traditional rankings, unclear entity signals, third-party corroboration or several of those things together. It may also be too early to identify the cause with confidence.

The burden is to explain why the proposed action makes more sense than those alternatives.

A change can be technically correct and still be a poor decision when the likely upside is weak, the evidence is thin or the collateral risk is high.

If an existing page already owns the subject, protecting that ownership may matter more than publishing something new. If a technical issue affects one retrievable layer but not the actual visibility problem, fixing it may be correct without being sufficient. If the available evidence points mostly outside the website, rewriting five pages may simply be the wrong investment.

The work needs to follow the diagnosis rather than the agency’s preferred deliverable.

Search case studies often get cleaned up after the result is known. Real search work is usually messier than the finished story makes it look.

Suppose an agency changes a group of pages in April, AI citations rise in May and organic traffic improves in June. That sequence is worth examining, but it does not automatically prove the April work caused every later improvement.

Search systems change. Competitors move. Models and retrieval systems change. Other websites mention the brand. Demand shifts, traditional rankings move and the site itself may have changed in other ways.

The stronger case study preserves enough context for the reader to understand what was actually measured. I would expect the baseline date, the intervention, the observation window and the exact result being claimed. If there is a useful comparison group or unchanged section, show it. If obvious confounding factors exist, acknowledge them.

The wording of the claim should match the evidence. “Visibility improved after the changes” is not the same statement as “the changes caused the improvement.”

We apply the same standard to our own work. If the stronger claim cannot be defended, the claim should be narrowed. A slightly less dramatic result is preferable to a cleaner story that the evidence cannot carry.

Not every company needs the same engagement. Some businesses have no reliable baseline at all, while others already have good SEO execution but nobody systematically observing AI discovery. Some have obvious technical and content weaknesses that need implementation. Others already have content, SEO, development, PR and analytics specialists but no clear decision rights across them.

The missing capability may be measurement, interpretation, execution or leadership. That should be clear before the engagement starts.

A vague goal such as “we need GEO” makes it easy to buy activity. A more useful objective is to understand where the brand currently appears, where it does not, which gaps are actually actionable, what evidence supports the recommendation and how the team will know whether anything changed.

Now there is something concrete to manage.

If you are still deciding what the service side should include, our AI search strategy and implementation page explains how we approach the work. If the real problem is that you do not yet have a trustworthy starting point, a search visibility baseline and audit is usually the more sensible first move.

After all of that, my decision is fairly simple. By the end of the evaluation, the measurement, the observed result, the preserved evidence and the point where judgment entered the process should all be visible. Just as important, I need to know which competing explanation was taken seriously, why the proposed action deserves to happen now, and what would cause the team to stop or change direction.

The longest tool list does not answer those questions. Proprietary vocabulary does not answer them either, and confidence is useful only when it is proportional to the evidence behind it.

There are parts of search where we know quite a lot and other parts where we are working from patterns, experience and incomplete evidence. I trust the agency that can tell the difference, make a decision when the evidence is strong enough and decline to manufacture certainty when it is not.

For me, choosing an AI SEO agency comes down to that: whether the reasoning still holds up after you start questioning it.

For more practical material on measurement, citations, AI discovery and modern search, browse the DankRank AI Search Resources.