Before You Hire an AI Search Agency Check Whether You Are Buying SEO Twice

An AI search agency should not replace competent SEO with a newer acronym. It should add a bounded capability that the existing search work does not observe, diagnose or deliver. If two proposals crawl the same pages, rewrite the same copy, discuss the same entities and report the same Search Console numbers, the second label has not created a second discipline. It has created duplicate billing.

Google makes the overlap difficult to ignore. Its guidance says the established SEO foundations for Google Search remain relevant to AI Overviews and AI Mode. Google also says those features require no special schema, new machine-readable file or additional technical requirement beyond normal Search eligibility (Google AI features guidance). That does not make specialist AI Search Optimisation pointless. It makes the buying decision more demanding.

Start with the failure, not the supplier species

We need GEO and competitors are winning AI are procurement slogans. They are not observable problems. Google AI Mode describes our service incorrectly, Bing cites an obsolete comparison, ChatGPT cannot access an important public page and our existing SEO supplier does not inspect answer surfaces give a provider something testable. Different failures authorise different work:

This table is triage, not diagnosis. An inaccurate answer may originate in a contradictory website, stale directory, external source or variable product behaviour. A missing citation may trace back to an ordinary crawl or index problem. The investigation must be allowed to return no website change. An agency paid only to produce tasks will always discover tasks. A useful diagnostic sometimes prevents them.

SEO still owns the foundations

A serious SEO scope should already connect technical access, indexation, demand, page ownership, content, authority and conversion evidence to a commercial decision. It should say who implements recommendations and who checks the production release. Do not let a fashionable proposal rename that foundation and charge to discover it again.

Google advises SEO buyers to request explanations and corroborating evidence, warns that nobody can guarantee a number-one ranking and does not endorse a service because the provider implies approval (Google guidance on hiring an SEO, Google guidance on third-party SEO services). Apply the same scepticism to AI-search guarantees. A newer interface does not make certainty easier to sell honestly.

Core work normally includes crawl and index diagnosis; query and customer-decision research; page ownership and information architecture; useful content and evidence; internal and external discovery; website design and technical implementation review; analytics and customer acquisition measurement; and validation after release. If the “AI service” repeats those lines, ask what product-specific work justifies the premium.

Specialist scope must name the product and observation

AI visibility is not one market, index, result type or measurable position. A specialist earns additional scope by preserving:

• Exact question or prompt
• Named product and mode
• Date, locale and relevant state
• Displayed claim, mention or citation
• Visible supporting source
• Public evidence available to correct or support it
• Observation limitation
• Page, profile or source decision resulting from the evidence

The product name belongs beside every finding. A report combining all answer systems into one percentage has hidden uncertainty rather than solved it. Google says AI Overviews and AI Mode retain standard Search eligibility and may use query fan-out across subtopics and sources (Google AI features guidance). OpenAI documents OAI-SearchBot as an access control relevant to inclusion in ChatGPT search summaries and snippets (OpenAI publisher FAQ). A crawl state or citation in one product therefore proves nothing about another.

Product-specific observation is a credible addition. Generic statements about entities, helpful content and schema are not automatically specialist value.

Native reports have raised the evidence bar

By September 2026, a credible proposal should account for evidence supplied by the platforms rather than pretending every AI observation requires a proprietary visibility dial. Google introduced a dedicated Generative AI performance report in Search Console in June 2026 and reported worldwide rollout by 31 August 2026. It covers impressions for property links in AI Overviews and AI Mode and can group results by page, country, date and device, subject to product thresholds and limits (Google Generative AI report). That is useful native evidence. It is not a prompt transcript, unlinked mention report, citation-quality judgement or sales-attribution model.

Bing’s AI Performance reporting covers supported Microsoft experiences and documents citations, cited pages and sampled grounding queries while warning that citation activity is not ranking, authority, clicks or engagement (Bing AI Performance). ChatGPT referral evidence may show a visit without showing every answer impression or explaining a later enquiry.

I would rather receive four honest partial measures than one immaculate fictional score. Digital growth decisions improve when uncertainty is visible enough to manage.

Put both providers against one scope sheet

Do not compare two polished decks. Make both suppliers complete the same work areas with their evidence, deliverable, owner and measurement limit.

Now the premium has somewhere to hide, or nowhere to hide. If the specialist adds product-level observation, external-source diagnosis and measurement capability absent from the current programme, the incremental value becomes visible. If both providers propose the same page audit and content rewrite, the business probably does not need two retainers. If neither owns implementation, source correction, observation cadence or the stop decision, you are buying recommendations designed to die in a backlog.

Ask for a sanitised deliverable, not another case study

Case studies show what a provider chooses to present. A working deliverable shows how the provider reasons when evidence is incomplete. Ask an SEO supplier for one sanitised diagnosis connecting a query or technical condition to a page, proposed action and validation step. Ask an AI-search specialist for one sanitised observation preserving the prompt, product, date, displayed claim, cited source, evidence limit and resulting page or source decision. The sample does not need client secrets. It must answer:

1. What was actually observed?
2. Why did that observation permit this action?
3. What would show that the action shipped, succeeded or should be reversed?

A cropped answer, prompt list or coloured gauge cannot answer those questions alone. This is where non-commodity work becomes inspectable. The provider demonstrates how an unstable observation becomes a bounded decision without inventing causation.

Choose among three honest buying models

There are three sensible structures.

I prefer the smallest buying model capable of owning the full decision. Adding suppliers increases handoffs, attribution arguments and duplicate recommendations. The extra expertise must outweigh that coordination cost. Do not add a retainer because the board learned a new acronym. Add one when missing work can be named, evidenced and assigned.

Red flags wear old clothes with new labels

Walk away from an AI-search pitch that:

• Guarantees mentions, citations or recommendations
• Claims private knowledge of ranking or model systems
• Sells special schema, an AI text file or forced formatting as a universal shortcut
• Reports one blended visibility score without product evidence
• Treats one prompt as a stable market sample
• Counts screenshots without inspecting their claims and sources
• Prescribes mass content before identifying a missing customer decision
• Cannot say who implements the proposed change

Apply equal pressure to an SEO agency that dismisses answer surfaces without investigating whether customers use them; reports rankings without page, enquiry or revenue context; mistakes publishing volume for demand coverage; cannot distinguish a technical finding from the business decision it affects; or uses AI terminology to repackage work already included in the retainer.

The shared red flag is an uninspectable claim. The provider wants the label to carry confidence the evidence cannot.

Price the incremental decision, not the acronym

Before comparing fees, separate baseline work from incremental work. For each line item, record:

• whether it already exists in the SEO, content, branding or analytics scope;
• which new product or evidence source it adds;
• which decision becomes possible because of it;
• who acts on the finding;
• how often observation is required;
• what artefact remains after the contract;
• what condition ends or reduces the work.

A premium may be justified for specialised observation, research, external-source reconciliation, technical access diagnosis or high-quality synthesis. It is not justified merely because familiar activities acquire AI prefixes. This protects AI marketing budgets from two opposite mistakes: buying theatre and dismissing genuinely new buyer behaviour. The scope sheet makes both errors harder.

There is also a conversion rate optimisation boundary. An AI-search provider may identify a new entry route or expectation, but the destination still has to help the visitor decide and act. If nobody owns that downstream experience, answer visibility becomes another attention metric detached from conversion.

Write the statement of work before choosing the label

Document the observed failure, affected product, evidence already available, pages or public sources that may change and the commercial decision the work must support. Then ask each provider:

• What do you inspect first?
• Which evidence confirms or rejects the suspected problem?
• Which proposed work is ordinary SEO?
• What additional AI-search capability receives the premium?
• Which artefacts remain after the engagement?
• Who implements and validates the recommendations?
• Which outcome is directly measurable, and which remains inference?
• When should this work stop?

Choose the provider whose answers survive the sheet. The correct outcome may be an SEO agency, a bounded AI-search specialist, one team genuinely capable of both, or no new supplier. That is not indecision. It is procurement doing its job.

The boundary is straightforward: genuinely emerging products can require new research and measurement before the business has perfect benchmarks. A provider does not need certainty to add value. It needs transparent method, bounded claims, useful artefacts and a stop rule. Make the scope earn the label before the label earns a fee.