TL;DR
- Published AI visibility ROI formulas fail before attribution is reached. They divide a return that accrues over years by an investment made in one quarter, which is a period mismatch, not a measurement problem.
- Whether spend is an asset or an expense is a rule, not a fact about the world. IAS 38.63 forbids you from recognising internally generated brand-type intangibles; IFRS 3 requires your acquirer to recognise the same thing at fair value. The asset exists — it just cannot exist while you own it.
- The UK moved R&D across that line in 2014 and measured national investment jumped without anything real changing. The test used to decide which side you are on is published, and parts of a visibility programme pass it.
- THE THREE LINES: a citation can only touch revenue, cost of sale, or cost avoided. The field claims the first, which is the one that needs a counterfactual and the one you will lose.
- Claim the acquisition-cost line instead. It is a portfolio ratio your finance team already computes, it needs no attribution, and it matches the economic thesis: being cited lowers the cost of the next win rather than conjuring a win from nothing.
- Every number in the model turns out to rest on the decay rate of an earned position — the one input nobody publishes, and the reason the corroboration record is the only part of the programme with a datable life.
1. The question finance is actually asking
The budget conversation has changed shape in 2026. Two years ago the question was whether to fund visibility work at all. Now it is funded, the invoices are recurring, and someone in finance wants to know what the return is — not rhetorically, but in the sense of wanting a figure they can put in a model next to the other figures.
The field has answers ready, and they arrive with a great deal of supporting statistics. The circulating formula is some version of AI-attributed conversions multiplied by average revenue per conversion, divided by total programme investment. Variants layer it: direct AI revenue plus weighted AI-assisted revenue plus weighted AI-influenced pipeline. Vendor decks published this year report $3.71 returned per dollar spent, or apply a 1.44 multiplier to attributed revenue on the grounds that AI-referred visitors convert 4.4 times better. One widely shared worked example takes a mid-market firm spending $10,000 a month, adds referral revenue to a brand-search lift to a conversion premium, arrives at $91,667 a month, and reports 817% ROI.
The standard critique of these numbers attacks the numerator: the attribution is soft, the layers double-count, the assumptions are generous. All true, and all beside the point I want to make here. Three of these formulas fail before attribution is reached — they fail on the arithmetic of the ratio itself, on the accounting status of the thing being valued, and on which line of the profit and loss account the benefit could possibly land in. Those three failures are fixable. The attribution argument mostly is not.
What is AI visibility ROI?
AI visibility ROI is the return generated by appearing in generative answers, set against the cost of the work that produced those appearances. It is unusually hard to state because the benefit is a change in a stock — how findable and how recommended you are — while ROI is a ratio defined over a flow, and the two only coincide when the asset is consumed inside the reporting period.
2. Asset or expense is a rule, not a fact
Start with the part that surprises most marketers, because it explains a great deal of the behaviour they find baffling in their own finance departments.
Paragraph 63 of IAS 38, the international standard on intangible assets, says that internally generated brands, mastheads, publishing titles, customer lists and items similar in substance shall not be recognised as intangible assets. Paragraph 64 gives the reason: expenditure on such items cannot be distinguished from the cost of developing the business as a whole. The standard also states plainly that expenditure on research, training, advertising and start-up activities will not create a recognisable asset.
Read that list again with a visibility programme in mind. A programme that works to hold a place in the source set for a category question is producing something extremely close to brand and customer-list equity, assembled from advertising-adjacent and research-adjacent spend. It is not near the boundary. It is squarely inside the prohibited set.
There is a second clause that closes the door behind you. Paragraph 71 provides that expenditure once recognised as an expense cannot later be recognised as part of the cost of an asset. You cannot run the programme for three years, observe that it worked, and then capitalise it retrospectively. Once expensed, always expensed.
And now the part that makes the whole thing feel like a joke at your expense. Under IFRS 3, an acquirer buying your company must recognise identifiable intangible assets at fair value, including exactly the brand-type assets you were forbidden to recognise yourself. Academic work in the Journal of Brand Management calls this an unacceptable dichotomy: forbidden when internally created, mandated when acquired. The asset is real enough to appear on a balance sheet. It simply cannot appear on yours.
The usual retort to all this is the famous figure that intangibles account for around 90% of the market value of large listed companies. Treat that number carefully, because it is a residual rather than a measurement: it is market capitalisation minus book value, so it contains brand, patents, workforce, expectations, momentum and error in one undifferentiated lump. It proves that the accounts do not contain most of what investors are paying for. It does not tell you what any particular programme is worth, and quoting it in a budget meeting invites the obvious reply that the same logic values every unmeasured thing the company does at whatever is left over. The residual is a reason to build a case, not a case.
KEY TAKEAWAY Your CFO is not undervaluing the work out of ignorance or malice. A standard requires the spend to be expensed, and a different standard requires a buyer to value the same thing as an asset. Arguing that visibility is an asset is arguing with a rule you will not win against — so make the argument on ground where the rule does not apply: which P&L line moves, by how much, and over what period.
3. The line has moved before
If the asset boundary were a fact about the world it could not be redrawn. It is a definition, and definitions get redrawn.
Research and development has been capitalised in the UK National Accounts since 2014, as part of the changes introduced by the 2008 System of National Accounts and the European System of Accounts 2010. Before 2014 R&D was classed as intermediate consumption — used up inside the production process, invisible in gross fixed capital formation. Afterwards it was investment, adding to the capital stock. Nothing about the underlying activity changed. The measured level of GDP rose by roughly 1.9% across the European Union on average and about 2.5% in the United States; in Ireland the capitalisation of R&D alone contributed between 3.7% and 4.1% of GDP.
Billions of pounds of national investment appeared because a committee changed a classification, in much the way that what a backlink is understood to be has been redefined more than once without the underlying object changing. That is the cleanest possible demonstration that asset-versus-expense is an accounting judgement about the durability and identifiability of a benefit, not a physical property of the spend.
Better still, the judgement has a published test. To count as R&D in the national accounts under the OECD’s Frascati Manual, an activity must be novel, creative, uncertain, systematic, and transferable or reproducible. Five criteria, all of them assessable by someone who is not an accountant, and none of them about the medium the work is published in.
4. The Frascati screen
Run your own programme through those five tests and something useful falls out — not an accounting position, but a portfolio diagnosis.
THE FRASCATI SCREEN
Score each workstream against the five criteria a statistical agency uses to separate investment from consumption.
Novel — does it aim at a finding that did not exist before you ran it? A first-party measurement does; a restatement of what is already ranking, or of what a deep research mode would produce unaided, does not.
Creative — is there a new concept or method, rather than an assembly of existing ones?
Uncertain — could it have come out the other way? If the conclusion was fixed before the work started, this is production, not investigation.
Systematic — is it planned, budgeted and repeatable at a steady, defensible pace, or is it a one-off sprint?
Transferable — can someone else use, reproduce or check the result, which is where a genuine authenticity premium separates from a claimed one? This is the criterion that earned placements satisfy and owned explainers usually do not.
A workstream scoring five is doing something a national statistician would classify as investment. A workstream scoring one is consumption with a content calendar attached.
Two honest caveats, because this instrument is easy to over-claim. First, passing the Frascati screen does not make the spend capitalisable under IAS 38 — marketing expenditure is excluded regardless of how novel it is, and taking this to your auditor as a capitalisation argument will waste an afternoon. Second, the screen is a design tool, not a scoring system to report upward: its value is that it tells you which parts of the programme are building something durable and which are refilling a bucket with a hole in it.
Applied honestly, most programmes discover that the majority of their spend fails four of the five tests, and that the minority passing them — original measurement, primary data, work that produces a result someone else can cite — is also the minority producing the placements that get cited back. That correlation is not a coincidence, and it is the first genuinely financial argument for the shape of a content budget rather than its size.
5. The period mismatch that voids the ratio
Return on investment is a ratio, and a ratio is only meaningful when its numerator and denominator describe the same period. This is not a subtlety; it is the first thing anyone is taught about the metric, and it is where every published AI visibility ROI formula breaks.
The spend is a run rate — a retainer, a team, a tool stack, month after month. The benefit is a position that persists after each month’s spend stops, decays at an unknown rate, and pays out across an unknown number of future periods. Divide one by the other and you get a number whose value depends entirely on where you drew the period boundary. An audit of where you are actually present will not rescue a ratio whose window was chosen for it. Take a quarter’s spend against a year of benefit and the ROI is spectacular. Take three years of spend against a quarter of benefit and the same programme looks like a disaster. Both figures are arithmetically correct and neither is information.
Worse, the period boundary is chosen by whoever is presenting, which makes the ratio a rhetorical instrument rather than a measurement one. An agency renewing a contract picks a window that starts after the programme ramped. An incoming CMO reviewing inherited spend picks one that starts at first invoice. Both are being honest about arithmetic and dishonest about inference, and neither is checkable by the person receiving the slide, because the window is rarely stated. This is the same defect that made early zero-click traffic modelling unfalsifiable, and the fix is identical: publish the boundary before you publish the number.
What is a decay rate for an earned placement?
A decay rate is the share of a benefit that disappears in a year if no further spend is applied. For an earned placement it compounds three separate risks: the page coming down or being rewritten, the page dropping out of an index, and the claim inside it being superseded by something newer. None of the three is exotic to measure — you need a list of your placements and a monthly check — and the number matters more than anything else in the model.
The business already knows how to handle this everywhere else. The median B2B software company recovers its customer acquisition cost in 16 months, according to the 2026 Aleph and Benchmarkit performance benchmarks drawn from full-year actuals across 342 companies, with the top quartile at six months or less and the bottom quartile beyond 24. Nobody looks at a 16-month payback and declares acquisition spend unprofitable in month three. Finance has a whole vocabulary for costs that pay back over multiple periods — payback, net present value, internal rate of return, cohort economics — and marketing keeps declining to use it.
Using it requires committing to two numbers in public, which is why the field avoids it. You must state a useful life for the benefit, and you must state a hurdle rate. Neither is optional and neither is unknowable. The Bank of England held Bank Rate at 3.75% on 30 July 2026 with the next decision due on 17 September, so the risk-free anchor is public; your finance team has a weighted cost of capital sitting above it and will hand you the number if you ask.
THE PERIOD TEST
Before any ROI figure leaves your building, state three things and show the arithmetic:
1. The annual benefit in pounds, on one named P&L line — not a blend of three.
2. The decay rate — what fraction of that benefit survives into the following year with no further spend. This is the number the entire model turns on and the one no vendor states.
3. The hurdle rate your business applies to any other multi-year commitment.
Then discount: NPV = Σ (annual benefit × survival^t) ÷ (1 + r)^t, less the discounted cost of maintaining the programme. Report NPV and payback, not a ratio. A ratio with no period attached is a number without units.
6. The three lines a citation can touch
There are exactly three places in a profit and loss account where being cited can show up. The field claims the first almost exclusively, which is unfortunate, because it is the only one of the three that requires you to prove something about deals you might not have won.
THE THREE LINES
What each line claims, the evidence it demands, who disputes it, and how it fares in a finance review.
| P&L line | What it claims | Evidence it demands | Survival in a finance review |
| Revenue | New business that would not exist without the citation | A counterfactual: what these buyers would have done otherwise. Cannot be observed, only estimated | Poor. Disputed by every other channel owner and by finance itself, because the claim is unfalsifiable at deal level |
| Cost of sale | The same revenue won for less selling effort | Two ratios you already compute, tracked over time within a constant segment | Strong. The inputs come from finance systems, not marketing ones |
| Cost avoided | Paid media you no longer need to buy for the same coverage | A spend line that actually fell, or a documented decision not to raise it | Good, but only if you really cut the budget. An avoided cost nobody avoided is a forecast |
The cost-avoided line, and why it is mostly fiction
The third line deserves a word, because it is the one most often claimed and least often true. Cost avoided means paid media you did not have to buy because organic coverage did the job. It is a legitimate line when the spend actually fell — a cancelled campaign, a reduced retainer, a bid cap you lowered and kept lowered. It is not a legitimate line when nothing changed and someone has computed what the equivalent clicks would have cost at auction prices.
That second version is the old traffic-value metric wearing a new suit, and it has always had the same flaw: it prices your position at the cost of a substitute you were never going to buy, in an auction whose prices are set by advertisers with different margins and different objectives. For example, a firm whose paid search budget rose 8% in the year it claimed £200,000 of avoided media cost has not avoided anything; it has computed a hypothetical. Finance spots this immediately, and the credibility lost on that line tends to take the other two down with it. The same scepticism applies to authority scores quoted as if they were prices.
The asymmetry is stark once it is laid out. The revenue line requires you to establish something about a world that did not happen. The cost-of-sale line requires you to divide one number your finance team already produces by another number your finance team already produces. Yet almost every deck built in the last eighteen months has gone for the revenue line, presumably because the number is bigger and the arithmetic is easier to present.
7. Claim the acquisition-cost line
Here is the recommendation this whole argument has been building toward, and it is deliberately modest: report the effect of your visibility programme on customer acquisition cost, and stop reporting attributed revenue.
Customer acquisition cost is fully loaded sales and marketing spend divided by new customers acquired. It is a portfolio ratio. It does not ask which deals to credit, because it does not credit deals at all — it divides a total by a count. That single property removes the entire argument that has consumed the field: no source field, no channel tag, no dispute about whose touch mattered. A programme that works shows up as a falling cost per win, and one that does not, does not — a cleaner test than reading across from a competitor backlink comparison.
This also matches the actual economic thesis behind every earned-coverage strategy, which the revenue framing gets wrong. The claim was never that an answer engine conjures demand from nothing. The claim is that buyers who arrive already having encountered you are cheaper to sell to — fewer meetings, shorter evaluations, less discounting to overcome unfamiliarity. That is a cost effect by construction. Reporting it as revenue misstates your own mechanism.
For example: a UK logistics software firm reporting to a private equity board replaced its AI-attributed revenue slide with a single cohort CAC chart for its mid-market segment. The number was smaller and less flattering. It also survived three consecutive quarterly reviews and a change of finance director, which the previous slide had not, because every input on it came from a system the board already trusted. That is the whole trade in one sentence: fewer pounds claimed, more pounds kept.
The confounding you must declare
CAC moves for many reasons that have nothing to do with visibility: a pricing change, a new sales hire ramping, a shift in mix toward larger accounts, a market that got easier or harder. Three controls keep the claim honest, and you should state all three in the same slide as the number.
- Hold the segment constant. Compare within one ACV band and one sales motion. Blended CAC across a self-serve tier and an enterprise team is not a measurement of anything.
- Use cohort CAC, not period CAC. Match the spend to the cohort it acquired, allowing for the sales cycle — currently averaging 134 days in B2B software and lengthening.
- Benchmark against your own trend, never against an industry median. Published CAC payback medians in 2026 range from under seven months to sixteen depending on which of four accepted formulas the compiler used. Your own prior year is the only comparison that holds method constant.
With those declared, the claim you are making is narrow and survivable: within this segment, over this period, the cost of acquiring a customer fell by this much, against a programme costing this much, and here are the other things that changed. A finance director can argue with that. They cannot dismiss it, which is more than can be said for 817%.
8. Worked example: Calderwood Compliance Software
Calderwood sells regulatory reporting software to UK building societies and credit unions from Edinburgh. £9.1M ARR, an average contract value of £74,000, and a sales cycle that has stretched past five months. Its board approved a £340,000 annual visibility programme in early 2025 — a research lead, two writers, an outreach retainer and a monitoring stack — and in January 2026 asked what it had bought.
The agency deck opened with 640% ROI. The CFO asked two questions and the number did not survive either: over what period, and which line of the P&L does the money arrive on? The rebuild took a fortnight and produced a considerably smaller number that has now been funded twice.
The arithmetic
- Cohort CAC, held to one segment. FY2024: fully loaded sales and marketing of £3.60M against 49 new customers, giving £73,500 per win. FY2025: £3.94M against 61 new customers, giving £64,600. ACV moved less than 2% and the sales motion did not change.
- The benefit, stated on one line. A £8,900 reduction in cost per win across 61 wins is £543,000 of acquisition cost avoided in the year. That is the entire claim. No attributed revenue, no assisted pipeline, no multiplier.
- The decay assumption, stated out loud. Calderwood assumed 25% annual decay on the benefit if spend stopped, and modelled a three-year life at a 12% hurdle rate.
- The discounting. Benefits of £543k, £407k and £305k discount to £485k, £324k and £217k, a present value of £1.03M. Three years of £340k programme cost discounts to £817k. Net present value: about £209,000, on a programme the deck had valued at roughly six times its own cost.
THE SENSITIVITY THAT MATTERED
At 25% annual decay the programme returns an NPV near £209,000. At 50% decay — a position that halves each year without new spend — the same benefit stream discounts to £860,000 and the NPV falls to roughly £43,000, close enough to zero that the decision becomes a judgement call.
Nothing else in the model has that leverage. Not the hurdle rate, not the CAC estimate, not the segment definition. The entire investment case rests on the persistence of an earned position, and that is the one input the industry does not publish.
That finding reorganised the programme. If decay is the swing variable, then the highest-value question is not how many citations were earned this quarter but how long each earned thing lasts. Calderwood began tracking the persistence of individual placements — whether the page stayed live, stayed indexed and stayed cited — and moved budget from formats with short observable lives toward ones that stay put. It also stopped commissioning anything resembling the patterns spam classifiers look for, or anything that could not survive a technical audit of its own crawlability, on the grounds that a placement no engine can reach has a decay rate of one.
9. Where this leaves earned links
Return to the accounting frame for a moment, because it produces the sharpest practical rule in this article.
An intangible is recognisable, in the standard’s language, when it is identifiable: separable — capable of being sold, transferred or licensed independently of the business — or arising from contractual or legal rights. That is the test an acquirer’s advisers apply during diligence, and it is a useful test to apply to your own programme long before anyone is buying you. Three questions do the work.
- Is it separable from you? Could it exist, and be checked, without your company continuing to operate?
- Does it survive a change of control? Would it still be there the day after an acquisition, without anyone renewing anything?
- Can a stranger verify it without your systems? Not your dashboard, your login, your export — a third party, from outside.
Now sort the assets of a typical AI visibility programme. Visibility scores fail all three: they are outputs of a vendor’s instrument, they stop the day the subscription stops, and nobody outside can reproduce them. Owned pages fail the first two: they are separable only along with the whole company, and they need someone to keep paying for hosting and maintenance. Prompt libraries, monitoring histories and the shared workspace pages holding them are working papers.
The corroboration record passes all three. A dated set of third-party publications naming you exists on other people’s servers, persists through a change of ownership without anyone renewing a licence, and can be verified by an acquirer’s advisers from a browser, with no signature or credential required. It is, in the strict sense the standards use, the only identifiable thing the programme produces — which is why it is also the only part that shows up in someone else’s fair-value exercise rather than disappearing into goodwill.
That gives buying criteria with a financial rather than an aesthetic basis. Prefer domains with institutional longevity over properties that may not exist in three years, and discount surfaces where attention is intense but brief. Prefer formats that stay live and stay indexed and reachable over ones that rotate off a homepage in a fortnight. Keep the register: publication, date, URL, what it says — the discipline that makes provenance records useful. It costs a spreadsheet, and it is the only line item a stranger can check.
10. The strongest objection
The best counter-argument is a practical one, and it lands. You have traded a large number for a small one. The CFO wanted a figure by Thursday and you have handed back a net present value with three declared assumptions, one of which you admit nobody knows. Meanwhile the ROI ratio, crude as it is, has the enormous advantage of being a shared convention: everyone in the room understands it, nobody has to be taught anything, and comparisons across programmes are at least superficially possible. And the CAC claim is contaminated by everything else the business did that year, which you concede yourself.
Most of that stands. The trade is real and it is not free. But the trade’s value depends on which conversation you are in, and the boundary is fairly precise. When a budget is being defended — a review, a cut round, a change of finance director — a small survivable number beats a large auditable one, because the large one invites exactly the audit it cannot pass. When a budget is being created — a new programme, a first hire, a category nobody funds yet — the modest number will not get you the meeting, and a directional case with stated assumptions is the honest version of the big story.
On contamination: this is a real limit and not a fatal one. A CAC movement is evidence about a portfolio, not proof about a channel, and it should be presented that way, with the other changes of the period named in the same breath. That is a weaker claim than the field is used to making. It is also a claim that has never, in my experience, been thrown out of a budget meeting.
The honest negative to finish with: run this properly and some programmes will show a negative net present value at a realistic decay rate. That is not a failure of the method. It is the method working, and far cheaper to discover in a spreadsheet than in year four.
11. What to do on Monday
- Stop reporting a ratio with no period attached. Any ROI figure needs a stated benefit life, a stated decay rate and a stated hurdle rate before it leaves the building.
- Ask finance for the hurdle rate the business applies to other multi-year commitments. It is one email and it changes every model you build afterwards.
- Pick one line. Report the acquisition-cost effect within a single segment, and drop attributed revenue from the deck entirely rather than reporting both.
- Rebuild CAC on a cohort basis, matching spend to the customers it acquired across the sales cycle rather than the calendar month.
- Write down your decay assumption and then go and measure it. Persistence of earned placements is tractable — track whether each one stays live, stays indexed and stays unaffected by content labelling changes — and nobody has published a good number.
- Run the Frascati screen across your workstreams and find out what share of the budget would count as investment rather than consumption — including the internal work such as link sculpting that rarely gets costed at all.
- Start the register today: every earned placement, with publication, date and URL. It is the only asset here that passes a separability test.
- Name the confounders in the same slide as the number. It costs you nothing and it is the difference between a claim and a pitch.
The uncomfortable conclusion of all this is that being cited will never appear on your balance sheet, and that the effort spent arguing it should has been effort wasted on a rule that will not move. What can be done is smaller and more durable: name the line, state the period, declare the decay, and keep the one record that a stranger can verify. Everything else in the programme is working papers, and working papers do not survive an acquisition, a change of finance director, or a bad quarter.
