Does an AEO agency write your content, or does your team?

An AEO agency may write your content, your team may write it, or the work may be split, and agencies in this category sell all three arrangements, and which one you are buying should be stated in the proposal rather than discovered in month two. The first is agency-drafted: the agency writes new pages and rewrites existing ones, your team reviews and approves. The second is brief-and-build: the agency produces the specification for each page, the questions it must answer, the structure, the metadata and the gaps, and your writers do the drafting. The third is rewrite-only: nothing new is published and the work is entirely on pages you already have.

Most engagements end up mixed, because the three jobs inside a page have different owners. Structure travels well and can be specified by anyone who knows how answers get retrieved. Domain facts do not travel at all: your engineer knows the tolerance, your operations lead knows the lead time, your finance team knows what the contract actually covers. An agency that writes those from the outside is guessing, and guessing is where this work goes wrong.

The rewrite-only model is more useful than it sounds and is often the right first quarter. Our own teardown of one of our own pages is that model applied to us: the page scored 28 out of 100 on our framework, the fixes took an afternoon, and the score reached 71 without a single new page being published. Whether to rewrite or publish in a given situation is covered in our guide on rewriting existing pages against publishing new ones.

What matters more than the model is that one person holds the pen on a page and one person owns the facts in it. Those two can be different people. They cannot be nobody.

Which parts of a page should an agency never write?

An agency should never write a number, a claim about your product, or anything a regulator or a customer could hold you to. Those are the parts with consequences attached, and an outside writer who fills them in is either reading them off something you published or inventing them.

The invented version has a measured cost now. Runze Kong and colleagues built LIVE-GEO, a benchmark that publishes rewritten product pages as live web pages and lets real engines retrieve and choose between them, covering 80 fictional brands across 18 product families. Their verifier flagged 86 of the 400 rewritten descriptions for numerical faithfulness problems, including all 80 of the descriptions produced by the strategy of adding statistics (arXiv 2611.01126, submitted 31 October 2026). The tactic most commonly sold as AEO, adding impressive figures to a page, was the one tactic that failed a check on whether its figures were supportable.

This is also where our own tooling draws its line, and it is a deliberate product decision rather than a limitation. Where a page is too thin to ground a fix, our scoring abstains and returns no fix rather than writing one, because a tool that scores pages on whether their claims are sourced cannot invent claims itself.

So the practical rule for a content brief is that every statistic on a page must come from somewhere a reader can check: your own measurement, a named third party with a date, or a customer who has agreed to be quoted. Where none of those exists, the honest page says less. A thinner page with four defensible claims survives scrutiny in a way that a fuller page with two invented ones does not, and the second kind is the sort a prospect catches in a sales call.

Does publishing more content get you cited more often?

Publishing more content is the weakest lever of the three available to you, and buying an agency to produce volume is the most common way this budget gets wasted. The evidence on what actually decides a citation points away from the page, let alone from the number of pages.

Mahe Chen and Nick Koudas analysed 158,099 generative engine answers carrying 468,037 citations and found that which pages an engine's own search retrieves explains about 32% of the variation in citations, while page properties explain 2.8%, with ChatGPT citing only pages its own search retrieves and naming brands it already knows even without a supporting page (arXiv 2610.15033, submitted 10 October 2026). Their summary line is the one to put in front of a volume proposal: "Pages are thus cited when searched; brands are also named when remembered."

Mike King of iPullRank, which sells services in this category, makes the operational version of the same point: "Publishing another article before investigating often leaves the original issue untouched" (4 November 2026). If your pages are not being retrieved, a twentieth page is a twentieth page that is not retrieved.

That does not make content pointless, and the distinction is worth being precise about. New pages earn their place when there is a buying question nobody on your site answers, because an answer that does not exist cannot be retrieved. They do not earn their place as a monthly quota. How much content is actually needed is covered in how much content you need to get cited by AI.

What should your team supply, and what should the agency supply?

Your team supplies the facts and the approvals; the agency supplies the structure, the research and the production capacity. Written down that way, the division is easy to hold, and it is also the division that survives a busy quarter when your subject experts have no time.

Model	Who drafts	Who supplies facts	Who approves	Fits when
Agency-drafted	Agency	Your experts, by interview	Your reviewer, per page	You have no writers and can give an hour per page to interviews
Brief and build	Your writers	Your experts	Agency checks against the brief	You have writers who know the product and need the specification
Rewrite only	Agency, on existing copy	Already on the page	Your reviewer, in batches	You have a library that nobody is reading
Hybrid, the common case	Agency for structure, you for claims	Your experts	Both, in one pass	Most B2B teams after the first quarter

The interview is the part buyers underestimate. An agency-drafted page about a technical product is only as good as the half hour someone inside your company spent explaining what is actually true, and the engagements that fail on content usually fail because that half hour never got scheduled.

Budget it explicitly: one subject expert, named, with a recurring slot. That single commitment separates the engagements where an agency writes usable pages from the ones where it writes competent pages about nothing in particular.

Should you let an agency publish AI-written pages on your site?

Let an agency publish AI-assisted pages where each one carries a real byline and claims you can defend, and draw the line at the point where the work becomes volume with nobody's name on it. Google's own position is that production method is not the test. Its guidance states that it focuses on "Rewarding high-quality content, however it is produced", while "Using automation, including AI, to generate content with the primary purpose of manipulating ranking in search results" is a spam policy violation (Google Search Central, 8 February 2023).

The spam policies themselves are the sharper boundary, and they reach the AI surfaces directly. Google's spam policy page defines scaled content abuse as "creating large amounts of unoriginal content that provides little to no value to users, no matter how it's created", lists "Using generative AI tools or other similar tools to generate many pages without adding value for users" as an example, and includes "attempting to manipulate generative AI responses in Google Search" in its definition of spam (read 12 November 2026).

There is now a second boundary, and it applies to the tooling rather than to Google. Anthropic's usage policy, effective 12 November 2026, prohibits using its models to "Manipulate the sources from which search engines or AI systems draw answers by seeding them with content" "that misrepresents its origin, authorship, or independence", and separately to conceal the sponsorship of content meant to influence decision-makers. Publishing AI-assisted articles on your own clearly branded site, as your company's view, is not what that describes. A network of apparently independent sites quoting each other is.

Two things to require of any vendor drafting with AI assistance: a real named author on the page, with a profile an engine can follow, and an owner for every factual claim. Whether AI-written pages get cited at all is covered separately in does AI-generated content get cited by AI.

What does the review process look like when the agency writes?

A workable review has three gates and takes less time than buyers expect, because each gate looks at a different thing. A factual gate, where someone who knows the product confirms every claim and number. A positioning gate, where someone confirms the page says what the company wants to be known for. And a publishing gate, where whoever owns the site confirms the page is structured and tagged as agreed.

Keep them separate and in that order. The common failure is a single review meeting where a marketing lead is asked to validate an engineering claim, approve positioning and check metadata at once, which produces an approval that has not really checked anything. The second common failure is a vendor shipping on silence after seven days, which is a reasonable contract clause and a terrible way to publish a technical claim.

Any vendor proposing automated changes should be able to answer King's question before you sign: "The review your team must do before any automated fix goes live" (iPullRank, 4 November 2026). If the answer is that no review is needed because the changes are structural, ask what happens when a structural change rewrites a sentence containing a regulated claim.

Where approvals are genuinely slow for compliance reasons, build the queue around it rather than fighting it: batch pages, pre-approve the claim library once, and let the agency work from approved language instead of asking for a new judgement on every page. Our guide on AEO when legal and compliance approve every page covers that pattern in detail.

How do you brief an agency so the pages do not read like everybody else's?

Brief an agency with the things only your company can say, because everything else on the page is available to your competitors from the same sources. A brief made of keywords produces a page made of consensus, and consensus is what every other page on that query already contains.

Three kinds of material do the work here. Your own measurements, even small ones: a benchmark you ran, a failure rate you track, a lead time you can stand behind. Your own experience of the problem, including the cases where the obvious approach does not work, which is the part buyers remember. And the questions your sales team hears repeatedly, in the phrasing customers use rather than the phrasing your product marketing prefers.

That last one is not a style preference. Alignment between the words on a page and the words in the question asked is the strongest page-level signal anyone has measured: Discovered Labs, analysing about 2 million AI citations across 10,000 crawled pages, reported prompt-content alignment at a standardised beta of +0.37 with a 95% confidence interval of +0.33 to +0.41, significant at q below 1e-73, and that "One standard deviation increase in alignment corresponds to roughly 30% more citations on average" (read 12 November 2026). Discovered Labs sells answer engine optimisation services, which is worth saying in the same breath as its finding, and in that study a feature had to survive at least four of nine robustness checks to be reported at all.

So the briefing question that produces a different page is not "what should we rank for". It is "what do we know that the twentieth article on this subject does not". If the answer is nothing, the page is not ready to be written, and a vendor who tells you so is worth more than one who writes it anyway. If you want to see which phrasings your competitors already own before commissioning anything, a free visibility assessment shows the raw answers and cited sources for your buying questions.

What else do people ask about who writes the content?

Marketing leads buying this work for the first time ask these five questions most often about content ownership.