TL;DR: Most of the work that earns AI citations happens on surfaces an agency cannot own. Review platforms need your customer relationships. Entity data needs product and legal sign-off. Community presence needs people who actually work at your company. An agency can own your content and your technical fixes, which is roughly a third of the citation surface, so the realistic model is a hybrid with an explicit split rather than a choice between the two.
Key Takeaways
Outsourcing has a hard boundary here. Of the six surfaces generative engines retrieve from, an agency can fully own one and partly own two. The rest require access or relationships that do not transfer.
The volume-versus-quality framing is wrong. Agencies are not the volume option and in-house is not the quality option. Retrieval selects passages, so volume past your question space hurts whoever publishes it.
Agencies win on pattern, you win on context. They have seen what works across many accounts. You have the sales calls, the product edge cases and the honest list of who your product does not suit.
Prompt sets should not be outsourced wholesale. The best prompts come from recorded sales conversations, which means the agency can format them but you have to supply them.
Hold agencies to rates with denominators. Cited Prompt Rate on category prompts, reported with its method. Mention counts and visibility scores are gameable by changing what gets tracked.
Neither option works without an internal owner. Someone has to approve entity facts, chase reviews and unblock engineering. If nobody holds that, the agency will stall and you will blame the agency.
Anchor targets assume your CMS generates slug IDs from heading text, which WordPress, Webflow and most headless setups do by default. Confirm on staging and adjust the hrefs if your platform slugifies differently.
Why this decision differs for GEO
In traditional SEO, an agency can own most of the work. They can research keywords, write and publish content, build links and run technical audits, and the majority of ranking factors sit on your domain where their access reaches. Generative engine optimization breaks that assumption, because the sources engines retrieve from are mostly not your domain.
The structural difference
Traditional ranking is dominated by pages you control. Generative answers are assembled from whatever the engine retrieved, which frequently means review platforms, third-party roundups, community threads and structured reference data. An agency can write your pages. It cannot get your customers to leave reviews, cannot approve your pricing facts, and cannot credibly post in your product's community.
That is the whole argument, and it reframes the question. You are not choosing who does the work. You are choosing who does the third of it that can be delegated, and deciding who owns the rest.
What the old framing gets wrong
The usual version of this comparison says agencies bring speed and volume while in-house brings quality and brand consistency. Both halves are misleading in this context.
Volume is not an agency advantage here. Retrieval selects passages, not pages, so near-duplicate content competes with itself regardless of who produced it. An agency publishing at high volume into an already-covered question space creates the same cannibalisation an in-house team would.
Quality is not an in-house guarantee. Plenty of internal teams publish pages that re-explain the category and carry no constraint statements. Brand consistency and extractability are different properties, and only the second one earns citations.
The real difference is access and knowledge, which is what the next two frameworks cover.
What each side genuinely brings
Agency | In-house | |
|---|---|---|
Pattern knowledge across accounts | Strong | Weak |
Speed to first measurement | Strong, they have the tooling | Slower, you buy and learn it |
Product and edge-case knowledge | Weak | Strong |
Access to customers for reviews | None | Strong |
Authority to approve entity facts | None | Strong |
Ability to unblock engineering | Indirect | Direct |
Credible community presence | None | Possible |
Continuity of measurement history | At risk on termination | Yours |
The bottom row deserves a note. If your agency owns the tracking subscription, your baseline and trend data live in their account. Ending the relationship can mean losing the measurement history that proves the programme worked, which is a specific version of the lock-in problem covered in our procurement guide.
Framework 1: The Outsourcing Boundary
The Outsourcing Boundary maps each of the six surfaces generative engines retrieve from against whether an agency can own it, share it, or cannot touch it. It exists because the agency conversation usually happens at the level of deliverables, and the useful level is access. An agency can produce anything it has access to produce, and the citation surface is mostly access you cannot delegate.
The six surfaces
Surface | Can an agency own it? | Why |
|---|---|---|
Owned domain content | Yes, fully | Writing and publishing transfer cleanly with CMS access |
Technical access and schema | Shared | They can specify and often implement, but your engineering controls deployment and CDN rules |
Third-party roundups | Shared | They can pitch and build relationships; some publishers want to hear from the brand |
Structured reference data | No | Wikidata, Crunchbase and directory profiles need factual authority and often legal sign-off |
Review platforms | No | Review requests go through your customer relationships. An agency soliciting on your behalf raises platform policy issues |
Community threads | No | Credible participation requires people who actually work at the company |
One of six is fully delegable. Two are shared. Three cannot leave your building.
What this means in practice
If your citation mix is weighted toward roundups and reviews, which is common for B2B SaaS, an agency can address a minority of it no matter how good they are. That is not a criticism of agencies. It is a reason to scope the engagement honestly rather than hiring one and expecting citation rate to move on its own.
The failure mode is predictable: you hire an agency, they do excellent work on your domain, your citation rate barely moves because the gap was in reviews and roundups, and the relationship ends badly with both sides feeling misled.
Worked boundary, illustrative
A B2B SaaS company audits its citation mix across 30 commercial prompts, then maps it against the boundary:
Surface | Share of citations | Agency can own | Who must act |
|---|---|---|---|
Third-party roundups | 29% | Partly | Comms plus agency |
Review platforms | 23% | No | Customer marketing |
Owned domain | 21% | Yes | Agency |
Community threads | 13% | No | Developer relations |
Structured reference | 10% | No | Product marketing |
Uncited model claims | 4% | No | Nobody, accept it |
An agency given full scope can move 21% outright and contribute to another 29%. The remaining 46% needs internal owners regardless of what you pay.
How to use this before signing
Run your own citation mix first, then show it to the agency and ask which rows they will own. An agency that claims all six is either misunderstanding the work or overselling. An agency that says plainly that reviews and entity data are yours, and offers to help you build the process, is the one that has done this before.
Framework 2: The Capability versus Context Trade
The Capability versus Context Trade sorts GEO tasks by whether they reward pattern knowledge across many accounts or deep knowledge of one product. Agencies have the first. You have the second. Assigning a task to the side that lacks what it needs is the most common reason hybrid arrangements underperform.
What each kind of knowledge is
Capability is pattern recognition built from running the same work across many brands. An agency that has fixed crawler access forty times knows where the blocks hide. That knowledge genuinely does not exist inside a company doing this once.
Context is knowledge of your product, your buyers and your edge cases. Which integration actually breaks at scale. What the three objections in every sales call are. Which customer segment your product genuinely fails. None of this is in your documentation, and an agency cannot acquire it quickly at any price.
Task by task
Task | Rewards | Assign to |
|---|---|---|
Crawler access diagnosis | Capability | Agency |
Schema implementation and validation | Capability | Agency |
Competitive citation analysis | Capability | Agency |
Tool selection and measurement setup | Capability | Agency |
Content formatting for extraction | Capability | Agency |
Prompt set construction | Context | You, agency formats it |
Constraint and non-fit statements | Context | You |
Category label and positioning | Context | You |
Pricing and capability facts | Context | You |
Comparison page claims about rivals | Context | You, with legal |
Review outreach | Context plus relationships | You |
Roundup pitching | Both | Shared |
The two tasks worth arguing about
Prompt set construction. Agencies will offer to build this and it is the task most often outsourced badly. The best prompts come from recorded sales conversations, which means the raw material is yours. The workable split: you supply twenty call transcripts and the three objections you hear most, they turn that into a tagged prompt set. That takes an hour of your time and produces something far better than keyword-derived prompts.
Constraint statements. Writing who your product does not suit requires knowing it and being willing to publish it. An agency will not volunteer this because it feels like criticising the client. Yet it is one of the shapes recommendation prompts reward most. If you outsource content without supplying this, you get pages that claim universal fit and lose constrained prompts.
The practical rule
Agencies should own anything where the answer is the same across many companies. You should own anything where the answer is specific to yours. When a task sits in between, the agency does the structure and you supply the substance, which is a workable division as long as somebody internal is actually available to supply it.
Framework 3: The Hybrid Split
The Hybrid Split is the division that works in practice, with obligations written in both directions. It exists because most GEO engagements fail on the client side rather than the agency side: the agency delivers, the client does not supply what only they can supply, and both parties experience it as the other one underperforming.
The split
Agency owns | You own | Shared |
|---|---|---|
Crawler and rendering diagnosis | Canonical fact sheet and category label | Roundup pitching |
Schema specification and validation | Review requests and recency | Comparison page claims |
Content formatting and extraction | Constraint and non-fit statements | Prompt set, you supply, they structure |
Measurement setup and reporting | Entity data across third-party profiles | Technical deployment |
Competitive citation analysis | Community participation | Quarterly strategy review |
Weekly triage facilitation | Engineering prioritisation |
What the agency owes you
Write these into the engagement rather than assuming them:
A named weekly triage, 45 minutes, where lost prompts are routed to owners rather than discussed
Full answer text and every cited URL for lost prompts, not a summary dashboard
Citation rate reported by prompt intent type, with the method stated
Their own manual spot-check match rate against the tool's data, reported quarterly
An explicit list of findings that require your action, with dates, so stalls are visible
Measurement data exportable to you, in your account or delivered quarterly
What you owe the agency
This half is the one that gets skipped, and it is why engagements stall:
A named internal owner with authority to approve facts, available weekly
Twenty recorded sales calls or transcripts in the first month
Sign-off on the category label within two weeks, from product and sales
Access to whoever can change CDN, WAF and robots rules, with an agreed SLA
A customer marketing contact who will actually run review requests
A decision on constraint statements, including willingness to publish non-fit
The clause worth adding
Measurement continuity. Specify that tracking runs in your account, or that full historical data is delivered to you quarterly in a machine-readable format and on termination. Without it, ending the relationship loses the baseline that proves the programme worked, and you start the next engagement with no history.
The honest test of fit
If you cannot commit to the second list, an agency will not fix your GEO programme. That is a resourcing problem rather than a vendor problem, and hiring an agency to work around an absent internal owner reliably produces a year of good deliverables and a flat citation rate.
Conversely, if an agency will not commit to the first list, particularly the full answer text and the exportable data, you are buying reporting rather than a programme.
What to ask an agency
Ten questions, sent in writing before the pitch. The answers separate agencies who have run GEO programmes from agencies who have added the term to an SEO deck. As with vendor selection, the reframing a question receives is often more informative than the answer.
The questions
# | Question | A good answer | A warning sign |
|---|---|---|---|
1 | Which of the six citation surfaces will you own, and which are ours? | Names reviews and entity data as ours | Claims to own all of it |
2 | How do you build a prompt set for a new client? | Asks for sales call recordings | Derives it from keyword volume |
3 | Will you show me full answer text and cited URLs for lost prompts? | Yes, as standard | Offers a visibility score instead |
4 | How do you verify your tracking tool's accuracy? | Manual spot checks, with a match rate | Has not considered it |
5 | What do you do when the client blocks progress? | Escalation path and visible stall log | Avoids the question |
6 | Does tracking run in our account or yours? | Ours, or data delivered quarterly | Theirs, with no export commitment |
7 | How do you report, and on what denominator? | Cited Prompt Rate by intent, method stated | Mentions and traffic |
8 | What is your position on publishing volume? | Mapped to question space, with a ceiling | A monthly article quota |
9 | Who writes our constraint and non-fit statements? | You do, we will prompt for them | We handle all copy |
10 | What would make you tell us to stop spending? | A specific answer | Nothing |
Question 10 is the one that matters
An agency that can name a condition under which they would recommend reducing scope is an agency thinking about your outcome. Reasonable answers exist: your buyers do not use AI assistants, your citation rate is already strong and the remaining gap is off-domain, or your technical foundation needs fixing first and that is not their work.
An agency with no such condition has a commercial model that only grows, which is fine to know and worth pricing in.
What a pitch should contain
Ask for a sample of real work rather than a methodology deck: one anonymised lost prompt, the full answer the engine produced, the URLs it cited, and what they recommended doing about it. Fifteen minutes with that artefact tells you more than an hour of slides, and agencies who do the work have it readily available.
On white-label reporting
If you are an agency reading this, white-label reporting is a legitimate and valuable feature. The caution is that a report branded as yours still has to survive a client asking what the denominator is. Reporting mention counts under your own logo does not make them defensible, and the client who eventually asks will be the one deciding your renewal.
Step by step: making the decision
Nine steps, about three weeks. The first four happen before you speak to any agency, because the single biggest determinant of whether an engagement works is whether you know your own citation mix going in.
Step 1: Run your citation mix. Twenty category prompts, five runs, three engines. Record every cited URL and bucket it across the six surfaces. This is the artefact the whole decision rests on and it costs an afternoon.
Step 2: Map your mix against the Outsourcing Boundary. Calculate what share of your citations sit on surfaces an agency can own. If that number is under a third, an agency alone will not move your programme, and you need internal owners first.
Step 3: Identify your internal owner. One named person with authority to approve facts and chase other teams, available weekly. If you cannot name them, stop here. Neither option works without this and an agency will not substitute for it.
Step 4: Decide what you can supply. Go through the what-you-owe list in Framework 3 honestly. Sales call access, category label sign-off, engineering SLA, customer marketing contact. Each one you cannot supply narrows what any agency can achieve.
Step 5: Choose the model. Fully in-house if you have capacity and the surfaces are mostly yours. Hybrid if you need capability you lack and can supply context. Agency-led only if you are starting from nothing and need speed to first measurement, and even then expect to take surfaces back within two quarters.
Step 6: Send the ten questions in writing. Compare written answers rather than pitch impressions. Reframing is information.
Step 7: Negotiate measurement continuity. Tracking in your account, or quarterly data delivery and a termination export. This is the clause nobody thinks about until they want to leave.
Step 8: Write the split down. Both lists from Framework 3, in the statement of work. Obligations in both directions, with the weekly triage named and scheduled.
Step 9: Review at day 90 against rates, not activity. Cited Prompt Rate by intent type, compared with your day-zero baseline. Activity reports are not outcomes.
The step that decides everything
Step 3. An agency with a responsive internal owner outperforms an in-house team with none. A brilliant agency with no internal owner delivers excellent work into a vacuum, and both sides spend a year frustrated.
If you are deciding quickly
Run steps 1 and 2. If over half your citations sit on surfaces an agency cannot own, the answer is to resource internally first and consider an agency for the technical and content layer afterwards. That takes an afternoon and prevents the most expensive version of this mistake.
Holding an agency accountable
The reporting problem in agency GEO work is that the easiest metrics to report are the easiest to inflate. Mention counts rise when the tracked prompt set grows. Visibility scores have no external definition. Activity reports measure effort rather than outcome. Specify the metrics in the contract so this is settled before the first report rather than argued about at renewal.
What to require
Metric | Definition | Why it resists gaming |
|---|---|---|
Cited Prompt Rate by intent | Cited prompts over tracked prompts, split four ways | Adding prompts does not inflate a rate |
Share of Answer | Your mentions over all vendor mentions, same prompts | Competitive, so it cannot rise just from your activity |
Prompt set composition | Count by intent type, branded excluded | Makes set-stacking visible |
Surface progress | Status against each owner's SLA, both sides | Shows whether stalls are theirs or yours |
Spot-check match rate | Their tool against manual checks, 20 prompts | Catches measurement drift |
The two gaming patterns to watch for
Prompt set stacking. Adding prompts you already win, or branded prompts where your name is in the question, lifts every aggregate number without any work. Require the prompt set composition in every report and exclude branded prompts from headline figures.
Denominator drift. The tracked set quietly changes between reports, so period comparisons are not like for like. Require that any change to the prompt set is flagged and that the previous set is reported alongside for one cycle.
Neither pattern necessarily indicates bad faith. Both happen naturally when nobody specified the rules, which is the argument for specifying them.
The report that works
One page, monthly: the five metrics above, the prompt set composition, a list of findings requiring client action with dates, and a short note on what changed and why. Anyone in the room should be able to challenge the conclusion using the same page.
If the monthly report takes more than ten minutes to read, it is probably substituting volume for substance.
Reviewing at renewal
Compare Cited Prompt Rate on category prompts against the baseline you measured before the engagement started. If you have no such baseline, you cannot evaluate the relationship and the renewal becomes a judgement about how the meetings felt, which is how agencies with weak results keep accounts and agencies with good results lose them.
Seven ways the relationship fails
Most of these are avoidable at contracting time. They are ordered by how often they appear rather than how dramatic they look.
1. No internal owner. The agency delivers findings that need facts approved, reviews chased and engineering unblocked, and nobody is available to do any of it. The work stalls and both sides blame the other. Fix: name the owner before signing, with weekly availability.
2. Scope that ignores the Outsourcing Boundary. The engagement promises citation improvement while covering only surfaces that produce a third of your citations. Fix: run your citation mix first and scope against it explicitly.
3. Prompt sets built from keyword exports. Fast to produce, and they miss the recommendation queries where decisions happen. Fix: supply sales call transcripts in month one and require the prompt set to be built from them.
4. Reporting on counts. Mentions rise when the tracked set grows, which looks like progress and is not. Fix: contract for Cited Prompt Rate by intent with the prompt set composition in every report.
5. Tracking in the agency's account. Ending the relationship loses the measurement history that proves the programme worked. Fix: the measurement continuity clause. Your account, or quarterly data delivery plus a termination export.
6. Content that claims universal fit. Agencies avoid writing who your product does not suit, because it reads as criticising the client. Recommendation prompts reward exactly that. Fix: you write the constraint statements, or explicitly authorise them to.
7. Quarterly reviews on activity. Pages published, audits run, hours delivered. None of it answers whether you are cited more. Fix: review against a baseline you measured before the engagement began.
The thread
Six of the seven are contracting decisions rather than performance problems. The agency you choose matters less than whether the scope matches your actual citation surface and whether someone internal is available to do the half that cannot be delegated.
That is also why the most common failure is not a bad agency. It is a reasonable agency hired against a scope that could never have produced the outcome, by a client who had nobody free to supply what only they could supply.
30/60/90-day roadmap and selection checklist
This works whether you are hiring an agency for the first time or repairing an engagement that is not delivering. The first thirty days establish facts you need regardless of which path you choose.
Days 1 to 30: establish your own position
Week | Work | Exit criterion |
|---|---|---|
1 | Run your citation mix across 20 category prompts | Cited URLs bucketed across the six surfaces |
2 | Map the mix against the Outsourcing Boundary | The share an agency could own, as a number |
3 | Name the internal owner and confirm weekly availability | A name, not a team |
4 | Work through what you can supply from the Framework 3 list | An honest yes or no against each item |
Days 31 to 60: select or rescope
Week | Work | Exit criterion |
|---|---|---|
5 | Send the ten questions to shortlisted agencies in writing | Written answers, comparable |
6 | Request a real work sample: lost prompt, answer, citations, recommendation | One artefact per agency |
7 | Negotiate measurement continuity and the split | Both lists in the statement of work |
8 | Baseline before work starts: 20 prompts, 5 runs, 3 engines | A before-picture you own |
Days 61 to 90: operate and verify
Week | Work | Exit criterion |
|---|---|---|
9 | Weekly triage running, 45 minutes, symptom to surface to owner | Two sessions held |
10 | First monthly report checked against the contracted metrics | Denominators present, composition shown |
11 | Internal owner clears the first batch of client-side actions | No findings stalled over two weeks |
12 | Day-90 measurement against the day-60 baseline | A comparison using the same method |
Honest expectation setting: ninety days is enough to confirm the working relationship, not the outcome. Citation rates typically move between days 45 and 90, and the surfaces an agency cannot own move over quarters. Judge the first quarter on whether the split is working and whether findings are being actioned on both sides.
Selection checklist
Citation mix measured and bucketed across the six surfaces
Share of citations an agency could own calculated
Internal owner named, with weekly availability confirmed
Honest assessment completed of what you can supply
Ten questions sent in writing and answered in writing
Real work sample reviewed: lost prompt, full answer, cited URLs, recommendation
Agency named reviews and entity data as your responsibility, not theirs
Prompt set method confirmed as sales-call derived, not keyword derived
Full answer text and cited URLs contracted as standard reporting
Cited Prompt Rate by intent contracted, with method stated
Prompt set composition required in every report
Measurement continuity secured: your account or quarterly delivery plus termination export
Position on publishing volume agreed, mapped to question space rather than a quota
Constraint statement ownership decided
Weekly triage scheduled with both sides attending
Baseline measured before the engagement starts
Renewal criteria written down with the notice date calendared
FAQs
Should I hire an AI SEO agency or build in-house?
Usually a hybrid. An agency can fully own one of the six surfaces engines retrieve from and partly own two more. Reviews, entity data and community presence need your relationships and authority. Scope the agency to what it can actually control and resource the rest internally.
What can an agency not do for GEO?
It cannot request reviews from your customers, approve your pricing and product facts, create or correct your knowledge-graph entries, or participate credibly in your product's community. Those three surfaces frequently account for half or more of a B2B brand's citations.
How much does an AI SEO agency cost?
Retainers vary widely and the published figures are not comparable. The more useful number is your total cost including the internal hours you must supply, which the engagement does not remove. Budget for an internal owner spending several hours weekly regardless of what you pay.
What should I ask an agency before hiring them?
Which citation surfaces they will own, how they build a prompt set, whether they will show full answer text and cited URLs, how they verify their tracking tool's accuracy, whose account tracking runs in, and what would make them tell you to reduce spend.
How do I stop an agency inflating its reporting?
Contract for rates rather than counts. Require Cited Prompt Rate by intent type with the method stated, the prompt set composition in every report, and a flag whenever the tracked set changes. Exclude branded prompts from headline metrics.
Who owns the measurement data?
Settle this before signing. If tracking runs in the agency's account, ending the relationship can lose the history that proves the work succeeded. Require your own account, or quarterly machine-readable delivery plus an export window after termination.
Can a small in-house team run GEO alone?
Yes, if someone owns it. The technical diagnosis and tooling setup are where agencies add most, and both are one-off rather than ongoing. A small team with a named owner and a weekly triage habit will outperform a larger team with neither.
How long before an agency engagement shows results?
Crawler fixes can change retrieval within days. Citation rates typically move between days 45 and 90. Being recommended rather than merely cited depends on reviews and third-party coverage, which you control more than the agency does and which takes two quarters or more.
Deciding who runs your generative engine optimization
The decision looks like a choice between two options and is actually a scoping exercise. Measure your citation mix, work out what share sits on surfaces an agency can own, and resource the remainder internally whichever way you go. Teams that skip that step hire well and are disappointed anyway, because the gap was never in the work the agency was doing.
If you take one thing from this: name the internal owner first. An agency with a responsive counterpart outperforms one without, by a wide margin, regardless of which agency you pick.
When neither option is right yet
You cannot name an internal owner. Wait until you can. An agency will not substitute for this and the engagement will stall in month two.
Your buyers do not research with AI assistants. Ask your last twenty closed-won customers how they found you before resourcing this at all.
Your technical foundation is broken. If pages are not indexed or render client-side only, fix that first. It is cheap, it is shared work, and nothing downstream functions without it.
You have no baseline. Spend two weeks measuring before you hire anyone. Without it you cannot evaluate the engagement or defend the renewal.
Start with the citation mix
If you want the surface breakdown this decision rests on without two weeks of manual sampling, run a free audit with Blazly GEO. It reports citation rates across ChatGPT, Gemini, Perplexity, Claude and Grok, shows which URLs are cited instead of yours, and flags crawler rules blocking retrieval. Bucket those cited URLs across the six surfaces and you have the number that determines how much of this an agency could ever own.
Summary
Agencies bring pattern knowledge, tooling and speed to first measurement. You bring product depth, customer relationships and the authority to approve facts. The surfaces split along the same line: an agency can own your domain content and share technical and outreach work, while reviews, entity data and community stay with you. Scope to that boundary, write both sets of obligations into the engagement, secure your measurement data, and report rates with denominators rather than counts. Then judge the first ninety days on whether the split is working rather than on whether the numbers moved, because ninety days is too short for the outcome and exactly right for the relationship.