AI SEO Services: What to Check Before Paying an Agency

Quick Answer: Before paying for AI SEO services, ask for a baseline of buyer questions, saved answers and sources, a distinction between mentions and recommendations, and a realistic estimate of improvement and work. Separate the audit from ongoing changes and retests. Reject guaranteed rankings or claims of access to a platform's internal systems ^1,2,3.

Summary in a mind map

AI SEO Services: What to Check Before Paying an Agency
│
├─ Evidence before a fee
│   ├─ Get dated answers and their tested buyer questions
│   ├─ Ask for a realistic work estimate
│   └─ Keep initial access read-only
│
├─ Measure the right thing
│   ├─ Separate mentions from recommendations
│   └─ Define ambiguous results before scoring
│
├─ Price the actual work
│   ├─ Audit ends with evidence and a proposed plan
│   ├─ Ongoing work names edits, approval and retests
│   └─ Compare bids using the same requested format
│
└─ Reject unsupported claims
    ├─ No third-party view of internal ranking data
    └─ No guaranteed position in generated answers

Start with the agency's evidence

An eight-person chiropractic clinic in Columbus receives a proposal promising that AI answers will recommend the practice. The owner has a simpler purchasing question: what would arrive after payment? A credible scope should specify a baseline, the evidence behind it, the work proposed and a way to check whether that work was completed. Google's advice to site owners says an SEO audit should include realistic estimates of improvement and of the work involved. It also recommends asking how success will be measured ¹.

A sample deliverable with the business name removed is a useful first test. It should show the tested question, the answer as seen, the date of observation and the source cited by the answer. A screenshot without the question or date cannot be repeated. A summary score without the answer cannot be checked. These are procurement standards for this service, not a claim that any search platform requires a particular report format. NIST's voluntary measurement guidance calls for documentation of test sets, metrics and tools during evaluation ³.

The clinic should also ask what the agency needs to inspect. Google recommends granting only read access to Search Console for an initial SEO audit. This is a concrete boundary at the sales stage: a proposal can describe its diagnosis before anyone receives the ability to make site changes. The clinic can decide later whether to approve work that actually changes pages or settings ¹.

Save the buyer questions and answers

For the clinic, “Who is a chiropractor near me?” and “Which practice handles a particular condition?” are different buyer questions. The agency should state which type it proposes to test and why. That choice should be fixed before the results are scored. Otherwise the question set can change after a disappointing answer, leaving the owner with a favourable-looking chart that does not describe the same test. NIST's Measure 2.1 guidance calls for documented test sets, metrics and tool details in a voluntary AI evaluation process ³.

The proposal should show the exact answer text, with the question, platform and observation date attached. If a cited source appears in the answer, preserve the source as displayed. A later retest can then compare like with like. The method should say how many questions are included and whether the same questions will be used again. This does not turn a changing answer system into a deterministic ranking; it makes the agency's observation inspectable ³.

The agency should explain how it handles an answer that cites a directory rather than the clinic website. The answer may still show the business, but the source route differs. The report needs enough detail to tell the owner whether the proposed action concerns the business's own page, information held elsewhere or a weak result that cannot be attributed confidently. Google recommends evaluating an SEO's proposed advice against its official guidance, rather than accepting a third-party tool's recommendations without scrutiny ¹.

Separate mentions from recommendations

The clinic wants customers, not a tally of appearances in arbitrary text. A mention means the answer names the practice. A recommendation means the answer presents it as a suitable choice for the question. Those are report definitions the agency and buyer should agree before work starts. A single answer can contain a name in a list without endorsing it. The distinction prevents an easy metric from being sold as a stronger business outcome ³.

The classification rule should come with an answer passage for each result. These three excerpts are hypothetical, not observations about a real clinic. For the tested question “Which Columbus chiropractic practice offers evening appointments?”, Mention: “North Grove Chiropractic appears in the local directory.” The passage names the clinic but makes no choice. Recommendation: “For evening appointments, try North Grove Chiropractic; its listed hours fit that need.” The passage explicitly suggests the clinic for the tested need. Unclear: “North Grove Chiropractic is one option; confirm current hours before booking.” The passage names an option but withholds a clear endorsement. The report should keep the question, full answer and passage with each label; a reviewer should explain ambiguous cases. NIST calls for documented test sets, metrics and tools ³.

Do not accept a proprietary “visibility score” as the sole output. It may be a useful internal summary, but the owner needs its definition and the underlying observations. Google says third-party SEO tools do not have access to its internal ranking data. The agency can document what it observed; it cannot turn its dashboard into a private view of Google's internal systems ^1,2.

Check the proposed site work

After the baseline, the agency should identify the pages or information it wants to change and explain why. Google lists content or site-structure review, technical advice, content development and optimisation for generative AI among services an SEO may provide. That is a menu of possible work, not evidence that all items are required for this clinic. The proposal needs to connect each action to an observed gap and a realistic effort estimate ¹.

The agency should distinguish a diagnosis from an edit. For example, a report may say that a service page does not answer a tested buyer question. The follow-on work could be drafting an answer, reviewing it with the clinician and publishing it after approval. The initial report and the later change are different deliverables. The owner should know which one the quoted price buys and who signs off clinical statements. That approval path is a practical contract control, not a promise of a particular search result ¹.

The proposal should check clinic pages against Google Search Essentials and its spam policies, including any AI-generated copy. Google says AI-assisted content must meet those standards; publishing many near-duplicate pages mainly to manipulate Search violates its scaled-content spam policy ². The proposal should identify the page and passage at issue, the change proposed, and who will review clinical accuracy.

Technical work needs the same specificity. For Google's generative AI features, a page must be indexed and eligible for a Search snippet; meeting those conditions still does not guarantee that Google will serve it ². The agency should identify the affected clinic page and explain how it will check the page status. Structured data is not required for generative AI search, and Google says there is no special schema.org markup for it ². The agency may propose structured data for ordinary rich-result eligibility, but it should name that purpose instead of selling it as an AI prerequisite.

Decide between an audit and ongoing work

A one-time audit ends with an evidence package and a plan. It can include the test questions, captured baseline, classification rules, issues found, proposed changes, estimated effort and limits of the estimate. Google's hiring guidance supports realistic estimates rather than promises of first place. The owner can use the package to decide whether to commission implementation ¹.

Ongoing work has a different shape. It should name the approved edits, who makes them, the review point and the agreed retest. A monthly invoice should correspond to identified work or measurement, not merely the continued existence of a dashboard. This is a purchasing rule for the clinic. NIST's focus on documented tests and metrics offers a useful model for asking whether the measurement remains comparable between reporting periods ³.

Put the two scopes in separate rows of the agreement. Include acceptance criteria for the audit and for each later work phase. A buyer should be able to say, “We received the baseline and estimate,” or, “We approved and checked this change.” Without that separation, a small organisation can pay for a continuing service while still waiting for the first inspectable result.

In words, the tree branches like this: begin with buyer questions and saved answers; if the evidence is unclear, ask for proof. If it is clear, scope the audit, approve specific edits and retest the same questions. Review the new answers against the earlier record.

Treat guaranteed rankings as a warning

Google explicitly warns that no one can guarantee a number-one ranking in its Search results. It also cautions against claims of a special relationship with Google. Its guidance says third-party tools lack access to internal ranking data. An agency that promises a fixed placement in Google while hiding its method conflicts with these published cautions ^1,2.

The same purchasing discipline applies to a claim about ChatGPT answers, but the cited Google pages do not establish how ChatGPT ranks or selects businesses. Ask the agency to explain which platform it measured and what it actually observed. Do not let evidence about Google Search be presented as direct proof of a separate platform's internal method. No source in this article supports a guarantee about ChatGPT recommendations; the appropriate contract term is an observable deliverable, not a promised position.

The warning is strongest when the agency refuses to reveal its planned changes. Google tells site owners to seek explanations of what an SEO will do, and notes that the site owner remains responsible for the hired company's actions. The clinic should be able to review proposed content and site access before approving implementation. A secret method is hard to price, check or stop ¹.

What this checklist cannot decide

An evidence package can show whether the agency performed agreed tests and work. It cannot isolate every cause of a changing generated answer. The question, the observed answer and the platform may change; a before-and-after comparison therefore has limits. The contract should describe the observation method and report uncertainty plainly. NIST's documented test-set and metric approach supports this discipline, but it is voluntary guidance rather than a guarantee of attribution ³.

This article cannot set a universal fee for AI SEO services or promise that a particular edit will produce a recommendation. The three cited sources address audit expectations, limits of third-party access and documented measurement. They do not supply a market price or a causal formula for ChatGPT placement. A buyer should ask for a priced scope based on the actual site and proposed work, then compare that scope with other offers ^1,2,3.

What to do next

Ask each bidder for one page showing its proposed questions, saved-answer format, mention and recommendation definitions, realistic work estimate and first review point. Use the same requested format for every bid. For context, read whether ChatGPT recommends your business, how to check AI mentions, what generative engine optimisation means and why a competitor may appear instead.

If you want a first scoped next step, request the AI Answer Audit. Keep the request tied to the buyer questions and evidence you need to review.