In-house or agency for AEO

The short answer: Keep the accountability in-house and buy the work that needs breadth. Answer engine optimisation is three jobs with different economics: diagnosis across several engines, implementation inside your own stack, and corroboration from third parties. The middle one needs merge rights nobody outside your company should hold. The first one is bursty and is worth buying. Hiring an agency to own all three is how programmes stall.

That is the conclusion. The rest of this post is the reasoning, the evidence we can show for it, and a check you can run this week.

What companies appear to be doing

We have one point of reference on where this work sits, and it is weak evidence. A self-reported survey of several hundred marketing professionals, published in June 2026 by a firm that sells in this category, found that more than half of respondents had handed the work to the existing search team, about one in five were running it across search, content and engineering together, and only a small minority had outsourced it. On money, roughly two in five reported no dedicated budget for it, most of the rest were spending a minority share of their search budget, and about half expected to increase that spend within a year. We do not link it, because it is a vendor's own marketing survey of a self-selected panel, and its authors describe their own findings as directional rather than definitive. Treat it as a vendor estimate, not a measurement.

So read it as a weak market signal rather than a verdict, and note what it cannot tell you. A self-selected panel cannot say what the population does, and it cannot say whether outsourcing is rare because the category is young, because the deliverable is unfamiliar, or for some reason the survey never asked about. What we would actually want to know, how often an unowned programme fails silently on the afternoon a firewall rule changes and nobody re-runs anything, is not in that survey or in any source we can open.

The three jobs, and which one travels

Treat AEO as one blob and the staffing question has no answer. Split it and the answer falls out.

Diagnosis is bursty. You need a fixed prompt set, several engines, repeated runs, and somebody who can tell a retrieval failure from an identity failure. You need that hard for a few weeks, then not again for a quarter. Work shaped like that is expensive to hire for and cheap to buy. We have no survey that tells us how many companies split it that way.

Implementation does not travel. Fixing what the diagnosis found means a pull request, a CDN rule, a schema change, a redirect map. All of it needs commit rights, deploy access and a place in someone's sprint. An outside party can write the patch. Somebody inside has to merge it. In the engagements we have run, the findings that sit open longest are the ones with no named internal approver; we have not counted this across enough engagements to put a figure on it, which is the whole argument of The first 90 days of AEO work.

Corroboration is slow and relationship-bound. Getting a trade publication, a developer forum or an industry directory to carry an accurate version of your facts is a quarter of work, not a sprint. Agencies and PR firms genuinely help here, because the relationships are the asset and you cannot hire them in a month.

Notice that the three jobs have different clock speeds. A single monthly retainer covering all three hides that, and hidden clock speeds are how a programme looks busy while nothing ships. What an AEO budget actually buys breaks the same split out in money terms.

Half the diagnosis is now free, and in-house only

Here is the part that changed the maths in 2026, and that a lot of staffing debates have not caught up with.

Google now reports generative AI visibility inside Search Console. The generative AI performance report launched on 3 June 2026 and was rolled out to all sites worldwide on 31 August 2026. Google's help page for the report says it covers impressions from AI Overviews and AI Mode, grouped by pages, countries, devices and dates, and the launch post lists hourly through monthly granularity. The help page also documents an export button for the chart and table.

That data is first-party. It is attached to a verified property, which means your in-house team can see it and an agency can only see it if you hand over access. For Google surfaces, that is a useful AI visibility signal you already own at no extra cost. We have not priced it against every alternative, so we are not claiming it is the cheapest signal available.

Now the limits, as Google's own help page states them, because they define what buying anything is for. The metric is impressions. The dimensions it documents are pages, countries, devices and dates, with no query dimension among them, so you can see that a page surfaced in an AI answer and not which question produced it. Most data is assigned to the page's canonical URL rather than to a duplicate, so a page with duplicates has its performance rolled up somewhere you may not expect. The usual 1,000 row limitation from the Search performance report applies. As of today the Search Analytics API query reference lists no generative AI type or dimension, so we cannot say the report is pullable on a schedule; if you want it automated, check the API reference before you plan around it. And the report is Google's own surfaces only. It says nothing about ChatGPT, Perplexity, Claude or Copilot, which is not the same as saying you are invisible there.

So the honest division is this. Impressions, by page, on Google, are a free in-house job. Knowing which prompts name you, on which engines, with what claim attached, is a measurement programme somebody has to run. The sample audit report shows what that second half looks like when the evidence is kept rather than summarised into a score.

The question that settles it

Skip the capability comparison. Ask one thing instead: who merges the fix.

Run this check. Take the last three things anyone recommended for your site, whoever recommended them. Write down the date each one was proposed, the date it shipped, and the name of the person who approved the deploy. If the gap is weeks and the names are all internal, buy diagnosis and keep execution. If the gap is months and you cannot name an approver, no external contract will help, because the bottleneck is not knowledge. Hire or appoint the owner first, which is the argument in Who owns AEO inside a company.

Two more things worth asking before you sign anything. Ask where the prompt set lives and whether you keep it when the engagement ends, because an uncomparable baseline after twelve months is worse than no baseline. And ask which engines get run, how many times per prompt, and whether retrieval is toggled off on some runs. A supplier who answers those three questions precisely is telling you something real about method. One who answers with a score and a dashboard is not.

Where each arrangement breaks

In-house breaks on single-engine tunnel vision. A team with access to one chat assistant can end up optimising for its quirks, then be surprised when another engine names a competitor for the same question. Engines differ enough in retrieval and citation behaviour that we would not assume a result on one carries to the rest. It also breaks on nomenclature: an internal team that spent a year arguing about AEO vs GEO vs SEO rather than checking crawl access has lost that year.

Agency arrangements break on access and on sameness. Access, because the people who can fix the thing are not the people who found it. Sameness, because a template that produced results for one client gets applied to the next, and engine behaviour is specific enough that the template leaks. Ask to see the method, not the case study.

The hybrid breaks too, mostly on the handover. A report that lands with no named internal owner is a PDF, not a programme.

Honest limits

This post does not establish that the in-house split it recommends outperforms any other arrangement. We have no controlled comparison, only the engagements we have run, and we have not counted them in a way an auditor could reproduce. The staffing figures we gesture at come from a self-selected vendor survey we decline to cite as evidence, so treat the sentence about what companies do as unverified. Everything we say about the Search Console generative AI report is read off Google's own documentation for Google's surfaces, not tested by us, and it tells you nothing about how any other engine behaves. We also have no clean figure for how much of an improvement survives a model update, and no break-even headcount at which in-house measurement beats buying it, since that depends on how many engines you care about and how often you re-run.

FAQ

Is an agency worth it if we already have an SEO team? Often yes, for the first diagnosis and for third-party corroboration. Rarely for implementation, because your team already holds the deploy keys and an outside party will queue behind them anyway.

Can we do this entirely in-house? Yes, if someone owns it and has time to run the same prompts repeatedly. The free Google report covers Google's own AI surfaces and nothing else. The gap is coverage of the other engines and the discipline of repeated runs, both of which are time rather than talent.

What should we never outsource? The decision about what ships, the account access, and the prompt set itself. Those three are what let you compare this quarter to last.

How do we tell a real AEO practice from a renamed SEO one? Ask how many engines they run, how many times per prompt, and what they do about variance between identical runs. Vague answers to those questions are the tell.

Sources

• Google Search Central blog, introducing Search Generative AI performance reports in Search Console, 3 June 2026, with the worldwide rollout note dated 31 August 2026. Checked 12 October 2026: developers.google.com/search/blog/2026/06/gen-ai-performance-reports
• Google Search Console Help, generative AI performance report (Search), covering AI Overviews and AI Mode, the pages, countries, devices and dates dimensions, canonical URL attribution and the 1,000 row limitation. Checked 12 October 2026: support.google.com/webmasters/answer/16984139
• Google Search Console API, Search Analytics query reference. Cited here for what it does not list: as checked on 12 October 2026, its type and dimension values include no generative AI option: developers.google.com/webmaster-tools/v1/searchanalytics/query
• The staffing and budget shares described in the second section come from a self-reported survey run by a firm that sells in this category, published June 2026. It is a vendor estimate from a self-selected panel, so it is described here and deliberately not cited as evidence.