AI SEO Service: What to Expect and How to Choose a Provider

The market uses several overlapping labels:

• AI SEO;
• Answer Engine Optimisation (AEO);
• Generative Engine Optimisation (GEO);
• LLM SEO;
• AI search optimisation.

Some providers also use "AI SEO" to mean conventional SEO carried out with AI software. Notionmind is one current example of that usage. That is a different proposition.

When comparing providers, the label matters less than the actual service contract.

A credible provider should be able to tell you:

• which business problem it is addressing;
• which buyer questions or search journeys are in scope;
• which platforms or search surfaces matter;
• what it will assess before making recommendations;
• which changes it will make itself;
• which changes depend on your internal teams;
• how progress will be reported;
• how the engagement will be evaluated afterwards.

For a deeper explanation of how AI SEO fits with internal search strategy, see Kojable's AI SEO strategy guide.

A separate specialist becomes useful when commercially important AI-search work falls outside what your current team can do reliably.

Four common operating models are:

Before hiring another provider, ask whether your current team can already:

1. identify the AI-search problem that matters;
2. distinguish a material gap from a vanity visibility issue;
3. determine what action is justified;
4. execute or coordinate the required change;
5. report results using understandable metrics;
6. compare later observations with the starting point.

If the answer is yes, a separate AI SEO agency may add complexity rather than capability.

If the answer is no, specialist support may be justified.

The underlying baseline and representation methodology are covered separately in Kojable's AI representation guide.

The exact workstreams will vary, but the engagement should normally make the following stages visible.

1. Problem and scope definition

The provider should start by agreeing what problem is being solved.

Examples include:

• the company is absent from commercially relevant vendor-research questions;
• AI answers repeatedly describe an outdated positioning;
• important capabilities are omitted;
• competitors are framed more clearly;
• existing AI-search reporting cannot explain what deserves action;
• the company has conflicting or outdated information across public sources.

"Improve AI visibility" is usually too vague to act as a useful engagement objective.

The provider should specify:

• relevant buyer questions;
• category or topic scope;
• markets or languages;
• platforms or surfaces;
• competitors where relevant;
• exclusions.

2. Current-state assessment

The provider should document the starting point before recommending work.

The service page does not need to prescribe one universal baseline methodology. What matters commercially is that the baseline is defined, retained and suitable for later comparison.

Ask:

• What will be measured before implementation?
• How many questions, platforms or observations are included?
• Which conditions will be recorded?
• What will be retained for comparison later?

If the answer is simply "we'll show you your AI visibility score", ask how that score is constructed.

3. Diagnosis

The provider should explain why a particular gap deserves action.

A useful diagnosis connects:

observed problem → supporting evidence → commercial importance → justified intervention

The action should follow the diagnosis.

4. Implementation plan

Every recommendation should identify:

• what needs to change;
• why it matters;
• where the change belongs;
• how it should be implemented;
• who owns it;
• dependencies;
• review or approval requirements.

This is where an engagement becomes operational rather than advisory.

5. Execution where included

Some providers advise. Others execute.

Possible execution may include:

• technical SEO work;
• content updates;
• new pages where justified;
• documentation improvements;
• factual or entity consistency work;
• structured data where appropriate;
• supporting evidence assets;
• legitimate third-party outreach or correction;
• measurement setup.

Do not assume these are automatically included because the provider describes itself as "full service".

6. Reporting and verification

The engagement should explain how the team will know whether anything changed.

This does not require the provider to promise a particular outcome.

It requires the provider to define:

• what will be rechecked;
• when;
• under which conditions;
• which metrics will be used;
• which limitations apply;
• what happens if the expected movement does not appear.

A provider does not need to expose proprietary software or internal intellectual property.

It should, however, be able to explain enough of the method for a customer to understand what is being purchased and whether a reported result is meaningful.

At minimum, look for the following.

Business problem

The proposal should state the problem in commercial terms.

Weak:

Increase GEO visibility.

Better:

Establish whether the company is represented accurately in high-intent category and vendor-comparison questions, identify material gaps and implement agreed owned-site improvements.

Questions and surfaces in scope

A proposal should define the relevant query or buyer-question set and the platforms being assessed.

It does not need to test every possible prompt.

It does need a defensible boundary.

Deliverables

The proposal should state what you will actually receive.

Examples:

• current-state report;
• prioritised diagnosis;
• technical recommendations;
• content briefs;
• revised content;
• structured-data recommendations;
• source analysis;
• implementation work;
• reporting dashboard;
• retest report.

Avoid engagements where several of these are implied but none is contractually clear.

Exclusions

Useful proposals state what is not included.

For example:

• digital PR execution;
• technical deployment;
• content production;
• third-party profile management;
• legal review;
• development work.

Exclusions make responsibility clearer.

Responsibilities

Identify what belongs to:

• provider;
• SEO team;
• development team;
• content team;
• product marketing;
• PR/comms;
• legal/compliance;
• executive or subject-matter reviewers.

Methodology

The provider should describe:

• what is measured;
• how observations are gathered;
• how recommendations are prioritised;
• what metrics appear in reports;
• how later comparisons are made.

Dependencies

Examples include:

• analytics access;
• Search Console access;
• CMS access;
• developer support;
• subject-matter experts;
• brand positioning documentation;
• product evidence;
• review/approval capacity.

An aggressive timeline means little if the required dependencies cannot move.

Review and approval process

Agree:

• who approves changes;
• how claims are checked;
• whether drafts require legal/compliance review;
• how rejected recommendations are handled.

Retesting and end state

The proposal should state when and how the work will be reviewed.

It should not require the customer to invent the measurement method after implementation.

There are at least four delivery levels.

Recommendation

The provider identifies and prioritises an action.

Example:

Update the enterprise security page.

The customer still needs to determine how.

Implementation guidance

The provider explains what should change and how.

Example:

Add current deployment-scale evidence, supported compliance claims and implementation requirements to the enterprise security page, then connect it to the relevant product and comparison pages.

The customer or another provider implements it.

Asset preparation

The provider prepares some or all of the required asset.

Examples:

• content brief;
• page draft;
• FAQ copy;
• structured-data specification;
• evidence summary;
• source-correction brief.

Direct implementation

The provider actually publishes, deploys or carries out the change.

Examples:

• updates the CMS;
• deploys technical SEO changes;
• publishes approved pages;
• updates controlled public profiles.

The engagement should specify which delivery level applies to each workstream.

Ask explicitly:

• Who changes the website?
• Who writes or edits content?
• Who approves factual claims?
• Who handles external outreach?
• Who publishes?
• Who measures the result?
• Who owns follow-up work if nothing moves?

"Full service" is not a useful answer unless responsibilities are attached to it.

Potential workstreams include:

Technical SEO

Relevant when crawling, indexing, rendering, architecture or discoverability problems prevent useful content from performing properly.

Content improvement

Relevant when existing pages do not clearly answer commercially important questions or reflect current positioning and evidence.

New content

Relevant when a real information gap exists.

It should not be the default response to every visibility problem.

Structured data

Relevant where supported structured data accurately represents page content and serves an established search purpose.

Google's current generative-search guidance does not require special AI-specific schema, so a provider presenting "AI schema" as a universal requirement should be able to explain exactly what it means and why it is appropriate.

Entity and factual consistency

Relevant when company information is outdated, conflicting or unclear across controlled surfaces.

Evidence development

Relevant when a commercially important claim lacks useful supporting proof.

This might involve:

• case studies;
• research;
• documentation;
• product evidence;
• expert material.

Source analysis

Relevant when external sources repeatedly appear around an important buyer question.

For detailed citation interpretation, see AI citations.

Digital PR or third-party correction

Relevant only where there is a legitimate evidence or information opportunity.

A provider should not turn "get cited by AI" into a justification for manufactured mentions or covert promotion.

Monitoring and retesting

Relevant when the company needs to compare how answers or search outcomes move over time.

Detailed AI-representation measurement belongs in the AI representation methodology rather than being repeated here.

It should not force the buyer to reverse-engineer an unexplained score.

Useful reporting can include:

• questions or topic set assessed;
• platforms or surfaces covered;
• changes since the previous period;
• implementation completed;
• brand mentions where relevant;
• recommendations where relevant;
• citations where relevant;
• representation or factual accuracy;
• relevant search performance;
• AI-assistant referral traffic;
• limitations;
• next recommended actions.

The exact metrics will depend on the engagement.

The important procurement rule is:

Ask what every percentage counts.

For example, a provider may report "citation share" while another reports "responses with at least one citation". Those are not necessarily the same metric.

Similarly, two visibility products may calculate share of voice differently.

For deeper measurement methodology, use the dedicated Kojable resources on AI representation, AI citations and AI visibility and GEO tools.

Ask the following.

What was measured before the work?

If the case has no defined starting point, later improvement claims are difficult to interpret.

What changed?

Look for an implementation record.

Did the provider:

• change site content?
• fix technical issues?
• create new evidence?
• update controlled profiles?
• run PR?
• only start tracking?

Monitoring and intervention are different activities.

Were before-and-after conditions comparable?

Ask whether:

• the questions stayed comparable;
• the same platforms or modes were used;
• run counts were similar;
• the observation window was documented.

A screenshot from March and a different prompt in June is weak evidence of an intervention effect.

What exactly does the metric mean?

A claim such as:

AI visibility increased by 80%.

needs a definition.

Ask for:

• analytical unit;
• numerator;
• denominator;
• eligible population;
• timeframe.

Is the provider claiming correlation or causation?

A later improvement does not automatically prove that one provider action caused it.

Search indexes, third-party content, competitors, retrieval systems and model behaviour can also change.

A credible case study can still report:

After the intervention, the tested answers changed across repeated comparable checks.

It should be more cautious before claiming:

Our change caused the model to recommend the company.

Is the result repeatable?

One favourable answer can be real and still be unstable.

Look for repeated observations rather than a single screenshot where the provider is recommended.

Is a single case being presented as a universal benchmark?

Provider case studies can show that an outcome occurred.

They rarely establish that the same result or timeline will apply to every company.

Guaranteed citations or recommendations

A provider does not control ChatGPT, Gemini, Perplexity, Google or another third-party AI system.

OpenAI states that there is no way to guarantee top placement in ChatGPT Search.

Guaranteed permanent rankings

AI and search environments change continuously. A permanent placement promise should be treated sceptically.

Fixed universal timelines

A provider may have operational delivery timelines.

That is different from guaranteeing that third-party AI systems will reflect an intervention within a fixed number of days.

Unexplained proprietary scores

A proprietary score is not inherently bad.

An unexplained score is difficult to evaluate.

Ask what it counts.

Screenshots without methodology

Screenshots can demonstrate an observation.

They do not prove prevalence, stability or causality.

Tactic-first proposals

Be cautious when the provider recommends:

• 50 new pages;
• mandatory schema;
• mass Reddit posting;
• directory submissions;
• digital PR;
• review campaigns

before establishing what problem those tactics are meant to solve.

Hidden implementation ownership

If "full service" ends with a report and a task list for your team, understand that before purchasing.

One strategy for every AI platform

Different systems expose different search, retrieval and citation behaviours.

A provider making universal mechanism claims should be able to support them.

You may not need one when:

• your current SEO agency can already monitor and act on relevant AI-search requirements;
• your internal team can diagnose and implement the changes;
• the main requirement is software monitoring rather than additional services;
• no commercially meaningful AI-search problem has been identified;
• nobody has capacity to implement recommendations;
• the objective is only to increase a vanity visibility score;
• the business expects guaranteed recommendations from third-party systems;
• the organisation does not yet have stable positioning, product facts or evidence to publish.

Sometimes the right next step is not another agency.

It might be:

• improving the baseline;
• fixing existing SEO;
• updating product information;
• creating stronger evidence;
• giving the existing agency a clearer brief;
• buying monitoring software;
• assigning internal ownership.

A good AI SEO provider should be willing to say when it is not the required solution.

The best operating model may involve more than one of these.

A specialist provider should explain how its work fits the organisation rather than claiming every adjacent discipline needs to be replaced.

Kojable helps teams understand how AI represents the company, identify meaningful representation and evidence gaps, guide justified improvements and verify comparable answers afterwards.

That means an SEO or AEO provider can operate alongside Kojable.

For example:

• Kojable may identify an owned-site positioning gap;
• an SEO or content team may implement the required page changes;
• Kojable can then help assess whether comparable AI answers move.

The same model can apply to PR, documentation, product marketing or another workstream.

The purpose is not to make every task an "AI SEO" task.

It is to identify the actual representation problem, route the work to the right owner and verify what changes.