In-House vs Agency vs Hybrid: The 2027 Operating-Model Decision

TL;DR

— “In-house or agency?” is the wrong question because it is asked at the wrong unit. You do not make-or-buy link building — you make-or-buy each decision inside it, and those decisions have wildly different economics.

— Make-vs-buy has a settled theory (transaction-cost economics). The variable that decides it is asset specificity × coupling — not cost per link and not monthly volume, the two things every 2026 comparison guide measures.

— Plot each decision on the Governance Grid: own the specific-and-coupled cells (entity, corroboration, adjudication, brand reputation); buy the commodity cells (tooling, retrievability execution, prospecting); bond a partner for place-bound cells; keep coupled-but-generic execution close.

— Everyone who does this well is a hybrid — but the industry draws the internal line by cost-and-control feel, which puts it in the wrong place (it outsources the coupled work and insources the commodity).

— The specific danger is hold-up: let an agency own your entity identity and you have handed a lever on your brand’s AI visibility to a party you cannot cheaply replace. The same decision also changes quadrant per market.

The question everyone is asking is malformed

Open any 2026 guide to link building operating models and you meet the same fork: build an in-house team, hire an agency, or run a hybrid. The guides then reach for two numbers to settle it. The first is cost per link — the 2026 pricing data puts the average acceptable backlink at roughly £400 (about $509), up 45% since 2022, with 80.9% of practitioners expecting further rises. The second is volume: below about fifteen links a month, agencies win on unit cost; above it, a fully-utilised in-house team wins. Cross the threshold, the argument goes, and you insource.

That entire frame is a category error, and it is worth naming precisely because it is so nearly universal. It treats link building — or its 2027 successor, earning citations from answer engines — as a single homogeneous good you purchase by the unit, like printer paper, where the only question is whether you buy it or manufacture it in-house once your consumption is high enough. But the work is not one good. It is a bundle of very different transactions: defining who your brand is as an entity; engineering the technical retrievability of your pages; earning third-party corroboration; adjudicating which citation gaps are real and worth fixing; producing and formatting assets; operating monitoring tools. Pricing that bundle by the link and buying it by the unit guarantees that you mis-govern almost every transaction inside it — you will apply one make-or-buy verdict to a dozen decisions that each demand a different one.

You can see the confusion in the guides themselves. One influential 2026 comparison concludes that in-house teams should own “strategy, execution, and reporting,” with the agency brought in only to move faster. Another, equally confident, recommends the opposite hybrid: the agency leads strategy and the in-house team owns content production. Both call their answer “hybrid.” They draw the internal line in opposite places and neither says why theirs is right. That contradiction is the tell. When expert advice splits down the middle on where to cut, it is usually because everyone is cutting by feel — trading off control against cost — rather than by a principle that says which side of the line any given decision belongs on.

Make-vs-buy was never a firm-level decision

There is a principle, and it is not new. The question of what a firm should make and what it should buy is one of the most worked-over problems in economics, and the person who answered it — Oliver Williamson, building on Ronald Coase — won a Nobel for it. The answer is not “buy until you hit a volume threshold, then make.” The answer is that make-vs-buy is decided per transaction, and the variable that decides it is asset specificity: how specialised the assets involved are to this particular relationship, this particular brand, this particular place.

The logic runs like this. When a transaction is standardised and the assets are generic — many suppliers can perform it, and switching between them is cheap — the open market governs it well. Competition disciplines price and quality, and if a vendor underperforms you replace them. But when a transaction requires assets that are highly specific — knowledge, relationships, or investments that are worth far more inside this one relationship than anywhere else — the market fails, for a reason with a name: hold-up. Once a specific investment exists, the party holding it can extract rents, because you cannot costlessly switch away. The defence against hold-up is to bring the transaction inside the firm — to integrate it — so the specific asset sits under your own authority rather than a counterparty’s.

Two properties push toward integration, then, and both matter here. One is asset specificity. The other is coupling — whether the transaction can be cleanly separated from your other decisions or is entangled with them. A transaction you can wrap a clean contract around is buyable even if it is somewhat specific; a transaction whose outcome depends on being adjusted in lockstep with three other decisions cannot be handed across an organisational boundary without the coordination collapsing. Ranked search happened to be low on both axes, which is why an entire outsourcing industry grew up around it. Generative citation is high on both. That is the whole story of this article, and the rest is working out what it implies decision by decision.

It is worth seeing why generative citation scores high where ranked search scored low, because the 2026 data makes the case better than intuition can. The signals that now predict whether a model cites you are dominated by brand. In Ahrefs’ study of 75,000 brands, branded web mentions correlate with AI-Overview visibility at 0.664 against just 0.218 for backlinks — roughly a three-to-one gap — and earned media accounts for something like 82–89% of all AI citations. Domain authority, the metric a decade of outsourced link building was built to optimise, has watched its correlation with citation fall to around 0.18. A brand mention is far more brand-specific than a backlink: it is a statement, in third-party language, about who you are, which drags citation work up the specificity axis. And because a citation is won only when entity, corroboration, retrievability and freshness hold together at once, the decisions are coupled rather than additive, which drags it up the coupling axis too. The commercial stakes are not marginal either — AI-referred visitors convert at roughly 14.2% against 2.8% for traditional organic. Both axes moved together, and the outsourcing model built for the old position never re-priced itself.

The five kinds of specificity in citation work

Williamson catalogued the forms asset specificity takes. Mapping them onto the actual decisions inside a 2027 citation programme is what turns an abstract principle into an operating model. Five of them apply directly.

Read the table top to bottom and a pattern falls out. The first four rows are highly specific — to a place, to your brand, to your reputation, to a moment — and the market governs them badly. The last row is generic, and the market governs it well. A programme that buys or makes “link building” as one thing is forced to treat all five rows identically, which means at least four of them will be mis-governed no matter which way the single decision goes. Insource the lot and you are paying full salary to own commodity execution the market rents more cheaply; outsource the lot and you have put your reputation, your entity, and your local standing under the control of parties who bear none of the consequences. The correct move is to treat each row on its own terms — which is exactly what a grid, rather than a slider, lets you do.

The Governance Grid

Put the two variables on two axes. Horizontal: how brand-, place-, or reputation-specific is this decision (low to high). Vertical: how tightly is it coupled to your other specific decisions (low to high). Every decision inside your citation programme lands in one of four quadrants, and the quadrant tells you how to govern it. This is the instrument — not a scorecard of quality and not a pros-and-cons list, but a map of governance modes, one verdict per cell.

Take the quadrants in turn. The Buy cell (low specificity, low coupling) is where the open market genuinely serves you best. Operating a monitoring tool, running a competitor backlink analysis, executing schema and retrievability fixes, producing niche edits or guest posts at volume, formatting a page for featured-snippet and answer extraction — these are decomposable, standardised, and supplied more cheaply and often better by a specialist vendor than by a full-time hire. Switching costs are low; a clean contract fits around each. Buy them, and buy them as discrete services rather than bundled into a retainer that quietly absorbs the coupled work as well.

The Make cell (high specificity, high coupling) is the opposite. Your entity definition, the corroboration architecture that reinforces it, the measurement-and-adjudication call that reads the joint state and decides which gap to close — these are maximally specific to your brand and tightly coupled to one another. The market fails here for both reasons at once: the assets are too specific to contract cleanly, and the coordination across the coupling collapses the moment it crosses an organisational boundary. This cell must be owned, by a single accountable internal owner — the full-stack owner role the profession has been re-integrating precisely because these decisions cannot be handed off. You literally cannot buy accountability for your own brand’s reputation, because the party you would buy it from bears none of the consequences.

The Bond cell (high specificity, low coupling) is the one the binary framing has no language for. A presence node inside a market where you have no standing is highly site-specific — its value depends on being genuinely in that ecosystem — yet you can draw a clean interface around it: own citation coverage for that market. You cannot run it cheaply from headquarters (site specificity), but you need not fully employ it either, provided you can bond a partner whose incentives are aligned and who owns an outcome rather than filling a quota. A spot-market vendor is wrong here because a site-specific asset invites hold-up; a bonded, relational partnership is right. Finally, the Coordinate cell (low specificity, high coupling) is a genuine trap: work that is not special in itself but is entangled with your specific decisions, so an org boundary through the middle of it costs more in coordination than it saves in price. Keep it close — not because it is precious, but because separating it is expensive.

Why “hybrid” is right for the wrong reason

Notice what the grid has done. It has not produced “in-house” or “agency” as an answer. It has produced a distribution: some cells owned, some bought, some bonded, some kept close. In other words, the correct operating model is always a hybrid — not as a compromise that splits the difference between control and cost, but as the unavoidable consequence of the fact that your decisions live in different quadrants. Everyone who runs a citation programme well is a hybrid. There is no principled operating model that is purely in-house or purely agency, because no real programme has all its decisions in one quadrant.

So the industry’s instinct that “hybrid works best” is correct. It is correct for the wrong reason, and the wrong reason matters, because it makes practitioners draw the internal line in the wrong place. When you arrive at hybrid by trading off control against cost, you tend to keep the expensive, visible, strategic-sounding work in-house and send the laborious execution out — or, if you are cost-led, you keep cheap production in-house and buy the senior strategic brains. Either way you are cutting by how the work feels, not by where it sits on the grid. And the two lines rarely coincide.

The most common miscut is the one the guides recommend outright: outsource “link building execution” to an agency and keep “strategy” in-house. That sounds disciplined. It is backwards. A great deal of the execution — the tactics that actually earn citations — is coupled execution that must carry your specific entity naming and reinforce your specific corroboration, which lands it in the Coordinate or Make cells and argues for keeping it close. Meanwhile a good deal of what gets called “strategy” — audits, competitor mapping, tool-driven reporting — is decoupled evidence collection that sits squarely in Buy. Cut by feel and you insource the commodity while outsourcing the coupled work. The line looks sensible on an org chart and fails silently in production, exactly where the coupling breaks: earned mentions reinforcing a different entity naming than the site declares, a citable passage that never links back to the entity home, a freshness cadence disconnected from the outreach that would corroborate it.

The hold-up trap: what you cannot let an agency own

Hold-up is worth making concrete, because it is the sharpest reason the make-or-buy line is not a matter of taste. Suppose you let an agency own your entity and corroboration architecture. Over a year they accumulate the human-asset-specific knowledge of your brand’s citation identity — how it is named, which sources reinforce it, how its knowledge-graph presence is wired, what the models currently believe about it. That knowledge is worth a great deal inside your relationship and almost nothing outside it. Which means it is a specific asset, and it now sits with a counterparty.

You are now held up. If the agency raises its price or its service slips, your options are ugly: pay, or switch and watch your entity coherence — the thing brand mentions and corroboration spent a year building — fragment as the knowledge walks out the door. You have handed a lever on your brand’s AI visibility to a party whose interests diverge from yours (their incentive is retainer continuity and deliverable volume, not your entity’s coherence) and whom you cannot cheaply replace. This is not a hypothetical failure of a bad agency; it is the structural position of any external party holding a specific asset. The defence is the textbook one: internalise the specific investment. Own the entity and corroboration decisions so the knowledge lives inside your own authority.

The same logic explains why brand-reputation defence belongs in the Make cell without argument. When a model mis-states what your brand is, or a manual action or reputational hit threatens the very authority signals citations depend on, or you are fending off a negative-SEO attack on your entity, the party that must decide and act is the party that bears the consequence. An agency can execute the cleanup; it cannot be accountable for your reputation, because accountability is the one thing that cannot be transferred across a contract. It stays where the consequence lands, which is with you.

The same decision, different markets

One more turn of the screw, and it is the turn that makes the grid a real operating tool rather than a static picture. The quadrant a decision lands in is not fixed — it depends on the market. A decision that is Make at home can be Bond abroad, because the specificity that governs it changes with your standing in that market.

Take corroboration. In your home market you own it: you have the relationships, the language, the read on which local sources carry weight, so it is specific-and-coupled and it lives inside. Now look at a market where you have no presence. The corroboration decision there is still specific — more so, in fact, because international citation performance varies sharply by country and the models judge authority against local sources — but you cannot execute it from headquarters, and treating it as a translation problem rather than an authority problem produces fluent content that no local editor recognises and no model trusts. That decision has moved from Make to Bond: you need a partner genuinely inside that market’s ecosystem, one who can actually show up and be present locally, owning citation coverage for the market. The relevant fact here is the fragmentation: only about 11% of domains appear in both ChatGPT and Perplexity results, and coverage in one market says little about another. Your operating model is therefore not one grid but one grid per major market, and the cells shift as your presence does.

When the model says buy — and it often does

It would be easy to read all of this as “insource everything important, outsource the scraps,” and that reading is wrong. The grid is not a brief for building a large in-house empire. Over-integration — owning cells that belong in Buy — is its own expensive failure: you carry fixed salary and management overhead for commodity work the market supplies more cheaply, and you still may lack the per-market presence you cannot build in-house quickly. The honest, pro-outsourcing half of the model deserves its own hearing.

A great deal genuinely should be bought. Agencies and specialist vendors hold real advantages on the Buy layer: depth of execution on schema and retrievability, mature tooling, cross-client pattern knowledge, and the ability to scale production up and down without you carrying the payroll. The best current thinking on AI-search operating models lands on a division that maps cleanly onto the grid: agencies can collect the evidence; the company must own the interpretation. Monitoring platforms and audits tell you where you and your competitors are cited — that is buyable evidence — but the decision about which prompts represent commercial demand, which citation gaps are worth fixing, and which changes are meaningful is the coupled adjudication that stays inside. A well-run agency engagement, on this view, ends with assets your own team can operate without the agency. That is a Buy relationship drawn correctly: a clean interface, no specific asset stranded on the vendor’s side, no hold-up.

Two current findings sharpen where that interface should fall. The first is that AI visibility is not a single reading but a distribution — answers vary across runs, prompts and time, so the same query can cite you on Monday and omit you on Thursday. Measuring it once is close to useless; interpreting the spread is a judgement about your specific commercial prompts, which is why it belongs in the Make cell even though the raw monitoring that feeds it is bought. The second is the gap that opens when a brand buys visibility without ownership: teams that license a monitoring tool and stop there consistently see which prompts competitors own and rarely close the gap, because the tool reports the state without deciding what to do about it. The metric an agency relationship should be held to is not links delivered, nor an “AI visibility score” that climbs on the vendor’s own dashboard — it is cost per useful fix actually shipped, and “useful” is a coupled adjudication only the owner can make. Buy the collection; own the call. Draw the line anywhere else and you have bought a report and sold your judgement.

The right test at each cell is a single question: does doing this well require being accountable for your specific, coupled outcome — or just being good at a decomposable piece of work? If the latter, buy it, and do not apologise for buying it. The failure the grid is built to prevent is not “buying too much.” It is buying the wrong cells — the specific, coupled, place-bound ones — while paying agency rates to own commodity cells you should have rented.

The objection: isn’t the hard work where the specialists are?

Here is the strongest objection to all of this, and it is a good one. The deep expertise — the people who genuinely understand schema, entity engineering, retrieval behaviour, the tooling — lives at the agencies, because that is where the reps are. Your model insources the hardest, most technical work (entity, corroboration, adjudication) to a single internal owner and buys the “easy” commodity execution. Isn’t that exactly backwards? Shouldn’t you buy the senior specialist brains and staff your own people on the routine production?

The objection conflates two different things that happen to both get called “hard.” One is execution depth — knowing how to implement schema correctly, how to run a retrievability audit, how to operate the tools. That is a specialist capability, it is decomposable, and yes, the agencies are often better at it than you will be. Buy it. The other is accountable judgement over a coupled outcome — deciding which lever to move for your entity given the joint state of all of them, and being answerable for the result. That is not a deeper version of execution; it is a different kind of thing entirely. It requires authority over your coupled inputs (which an external party structurally lacks) and it requires bearing your reputational consequences (which only you bear). The in-house owner is not out-executing the agency on schema. They are making the coupled call, using the agency’s execution as an input. Depth-of-execution and accountability-of-judgement are different assets, and the theory says buy the first and make the second. The model does not insource the hard work to a generalist; it correctly identifies that the genuinely hard part is coordination and accountability — a governance function — not execution.

It is worth separating this from the familiar “core versus context” heuristic, because they look similar and diverge in a way that matters. Core-versus-context asks: is this differentiating? Transaction-cost reasoning asks: does market governance fail here? Usually they agree, but at the Bond cell they part company. A place-bound presence node may be entirely non-differentiating in the core/context sense — nothing proprietary about “having a credible contact in Germany” — yet it still cannot be left to a spot-market vendor, because its site specificity invites hold-up. Core/context would wave it out to a vendor and you would get burned. The grid keeps it as a bonded partnership. That is why the operating decision wants the specificity lens, not just the differentiation lens.

A worked reassignment

A UK direct-to-consumer brand approached its operating model as one decision: in-house or agency? It chose agency, on cost and speed, and because the agency arrived with slick AI-visibility tooling and a persuasive deck. Six to nine months in, the agency was shipping schema, running retrievability audits, producing content, and reporting a rising “AI visibility score.” Yet the brand was cited inconsistently, and in two quiet coupled failures the agency’s earned mentions reinforced a slightly different entity naming than the brand’s own product pages declared. The agency owned corroboration outreach; the brand’s product team owned the on-site entity; nobody owned the coupling across the contract boundary. Four plausible green dashboards, and the citations would not come.

The grid diagnosis was blunt. The brand had outsourced two Make cells — entity and corroboration — to a market vendor, retaining nothing but oversight, and was simultaneously paying agency retainer rates to own low-specificity retrievability plumbing that belonged in Buy. Everything was in the wrong quadrant at once: the specific-and-coupled work rented out, the commodity work over-owned. The fix did not touch how hard anyone was working. It moved the cells. Entity definition, corroboration architecture, and measurement-and-adjudication came in-house under one accountable owner. Retrievability execution, tooling, and content production kept being bought from the agency — but as discrete, decoupled services with clean interfaces, not a bundle. For the two markets where the brand had no standing, it bonded two in-market partners who owned citation coverage outright, rather than routing them through the UK agency. Citations stabilised within a couple of quarters. Notably, total agency spend fell — the brand had been paying retainer rates for coordination the agency was never structurally able to perform — even as in-house cost rose. The headline was not “in-house beat agency.” It was: the right governance for each decision, which is a hybrid, drawn by specificity rather than by feel.

What to do Monday

The five-step operating-model reset

1. Change the unit. Stop asking “in-house or agency?” List your actual decisions — entity definition, corroboration architecture, retrievability, adjudication, per-market presence, execution, tooling. That list, not “link building,” is what you make-or-buy.

2. Score each on two axes. For every decision, rate how brand-, place-, or reputation-specific it is, and how tightly it is coupled to your other specific decisions. Two quick highs/lows per decision is enough to place it on the grid.

3. Assign by quadrant. Specific-and-coupled: own it under one accountable owner. Commodity-and-decoupled: buy it as a discrete service. Specific-but-separable: bond a partner who owns an outcome. Generic-but-coupled: keep it close under tight coordination.

4. Run the hold-up test on everything you buy. Ask: if this vendor doubled its price or walked tomorrow, how exposed is my brand’s citation identity? High exposure means you mis-classified a Make cell as Buy — pull it in before the specific asset strands on their side.

5. Redo the grid per market. The same decision changes quadrant where you lack presence. Corroboration is Make at home and Bond in a market you don’t yet stand in. Your operating model is one grid per major market, not one grid for the company.

The 2027 operating-model decision is not a point on a slider between in-house and agency. It is a portfolio of governance choices, one per decision and per market, and the winning shape is a deliberate hybrid whose internal line is drawn by asset specificity and coupling rather than by cost per link or monthly volume. Get the line right and the familiar failure — clean dashboards, rising scores, no citations — stops being mysterious. It was never a quality problem. It was a governance problem: the wrong decisions in the wrong hands.