---
title: "Cutting a saturated lever: eBay, P&G and a pre-cut check"
description: "Is cutting spend once returns diminish a retreat? The eBay experiment, the P&G and Chase cuts, and a worst-case check to run before you cut, with limits."
url: "https://staging.khoi.co.uk/en/blog/cat-ngan-sach-don-bay-bao-hoa"
lang: en-GB
author: "Lê Quốc Khôi"
date: 2026-11-03
updated: 2026-11-03
topics: ["Strategy & leadership"]
tags: ["Saturated lever", "Theory of constraints", "Advertising effectiveness", "Field experiments", "Marginal returns"]
---

# Cutting a saturated lever is not retreat, if you check first

## Key takeaways

- eBay's field experiment: when paid ads on brand keywords were switched off on MSN, 99.5% of the clicks lost were recaptured by free organic search (Blake et al., 2014, p. 10). Non-brand ads switched off in about 30% of US market areas for 60 days gave a short-term return on investment (ROI) of -63%, 95% interval -124% to -3% (Blake et al., 2014, pp. 6, 7).
- Procter & Gamble (P&G) said it cut more than $100 million of digital advertising in the quarter to June 2017 and "didn't see a reduction in the growth rate" (Bruell and Terlep, 2017); JPMorgan Chase moved from roughly 400,000 sites to 5,000 pre-approved ones (Maheshwari, 2017). Both are company statements with no control group.
- Average return can be positive while the last pound spent loses money, because response to marketing is usually concave (Hanssens, 2024). Trimming differs from stopping: for 41 alcoholic drinks and mixer brands in Australia that went a year or more without advertising, sales fell on average by 16%, 25% and 36% after one, two and three years (summary only) (Hartnett et al., 2021).
- The six-step check at the end is my own synthesis, not any author's procedure.

On 27 July 2017, on its earnings call, the finance chief of Procter & Gamble said that after cutting part of its digital advertising the company "didn't see a reduction in the growth rate" (Bruell and Terlep, 2017). If you hold a budget line that keeps taking more money while sales stand still, you will ask whether a cut like that is a concession, and how you would know you had not cut the part that keeps the business alive.

## Is cutting a saturated lever really a retreat?

No, provided the line is not the system's constraint and the money you release has somewhere to go. It is a retreat if the spend is holding a position, such as new customers or awareness, and you give it up without putting anything in its place.

The [Theory of constraints](https://staging.khoi.co.uk/en/framework/theory-of-constraints) gives the first half of that answer. Goldratt defines a constraint as anything that limits a system from achieving higher performance against its goal (Goldratt, 2018, p. 3). For every other resource he asks whether there is any point in managing it to supply more than the constraints will consume, and says no: the performance of the whole system is already decided by the constraints (Goldratt, 2018, p. 4).

I call a line of spend that is near its target, where more money adds almost nothing, a [saturated lever](https://staging.khoi.co.uk/en/glossary#saturated-lever). If what limits your sales is something else, such as delivery capacity or the share of enquiries you close, more impressions will not lift the result. Applying this to advertising is my extension: the text I read speaks about systems in general and has no advertising-budget example. Goldratt also writes: "We very rarely find a company with a real market constraint, but rather, we find devastating marketing policy constraints" (Goldratt, 2018, p. 4). A rule such as "advertising is x% of revenue" could be one of these; that is my reading.

## What did eBay find when it switched ads off?

For a brand everyone already knew, it lost almost nothing in the short run, and most of the money went on people who would have bought anyway. Blake, Nosko and Tadelis ran two experiments at eBay and then calculated an ROI from the results.

| Test | Design | Result |
|:--|:--|:--|
| Brand keywords (queries containing "eBay"), 2012 | Paid ads halted on Yahoo! and MSN, kept on Google as the control | 99.5% of the clicks lost from the paid channel were recaptured by organic search (Blake et al., 2014, p. 10) |
| Non-brand keywords | Ads halted in 68 US market areas for 60 days; 142 areas left on | Ads added only 0.66% to sales, not statistically significant (Blake et al., 2014, pp. 14-15) |
| Short-term ROI | About $51 million a year of US paid-search spend, which the authors imputed from public data | -63%, 95% interval -124% to -3% (Blake et al., 2014, pp. 7, 23) |

What matters is who responded. New registrants and people who had bought only once or twice in the previous year bought more when shown ads; for people who bought more than three times a year the effect was close to zero (Blake et al., 2014, p. 17). Yet most paid-search traffic and credited sales came from frequent buyers (Blake et al., 2014, p. 18). The authors conclude that most of eBay's paid-search spending was wasted, because most of the clicks came from people who would have bought anyway (Blake et al., 2014, p. 19).

## What did P&G and Chase cut, and what do those cases not prove?

Both said the cut did no harm, but those are the companies' own reports with no control group, and most of what was cut looks more like poor placements than like a good lever that had saturated.

| | P&G (quarter to 30 June 2017) | JPMorgan Chase (March 2017) |
|:--|:--|:--|
| What was cut | Almost all of the quarter's advertising cuts were digital, aimed at placements that could end up on bot-driven sites or next to objectionable content (Bruell and Terlep, 2017). Ad Age, via Marketing Dive, put the fall at $140 million (Kirkpatrick, 2017); the Wall Street Journal (WSJ) said "more than $100 million" | Display ads from about 400,000 sites a month to 5,000 pre-approved ones; only 12,000 sites (3%) produced any activity beyond an impression (Maheshwari, 2017) |
| What they said about results | Chief financial officer Jon Moeller: "we didn't see a reduction in the growth rate"; organic revenue rose 2% (Bruell and Terlep, 2017) | Chief marketing officer Kristin Lemkau: "It's only been a few days, but we haven't seen any deterioration on our performance metrics" (Maheshwari, 2017) |
| What the source does not tell us | Sales had spending stayed the same. That year P&G also committed to more linear TV advertising, and described the restriction on digital spending as temporary and about brand safety (Kirkpatrick, 2017) | The outcome beyond a few days. The article says only that the cost of impressions and the visibility of the ads changed little, and names no sales or new-customer figure. Chase also pulled ads from YouTube the same week (Maheshwari, 2017) |

In November 2017 P&G's Marc Pritchard said that still hitting sales objectives after the cut showed "there is a lot of waste out there" (Tode, 2017). I treat both cases as suggestive, not causal: cutting need not hurt, but the reason needs a control group like eBay's.

## Why can average return look fine while the last pound loses?

Because the response of sales to spending is usually concave: the first pounds earn a lot, later ones earn less, then nothing. Hanssens writes that response is typically concave or S-shaped, so ROMI (return on marketing investment) depends on the spending level, and the return on the last unit can serve as a common yardstick (Hanssens, 2024). Lewis and Rao's model has the same zone of over-advertising, where marginal return is negative while average ROI is still positive (Lewis and Rao, 2015, section 2.1).

**Hypothetical example.** A small online shop selling coffee equipment spends £6,000 a month on search ads, at a 40% gross margin. Every figure below is invented by me; in real life nobody knows them, which is why you test.

| Slice of spend | Extra revenue | Extra gross profit | Net on this slice | Return |
|:--|--:|--:|--:|--:|
| First £2,000 | £12,000 | £4,800 | +£2,800 | +140% |
| Next £2,000 | £4,000 | £1,600 | -£400 | -20% |
| Last £2,000 | £1,000 | £400 | -£1,600 | -80% |
| All £6,000 | £17,000 | £6,800 | +£800 | +13% |

An average of +13% makes the line look healthy. But keeping only the first £2,000 yields £2,800 net, which is £2,000 a month better than keeping all £6,000. Cutting two-thirds here is not retreat: what goes is the loss-making part. The reverse holds too. If that first £2,000 is what feeds new customers, cutting it by mistake is the retreat.

## When is cutting a retreat after all?

When the spend buys something current sales cannot sustain on their own, or when you stop altogether instead of trimming. The sources give three warning signs.

First, new customers: at eBay the group that responded was the group that did not yet know the product (Blake et al., 2014, pp. 17-18). Second, size: Blake and colleagues write that their finding may not hold for "small and new entities that have no brand recognition" (Blake et al., 2014, p. 24). Third, duration: Hartnett and colleagues studied 41 Australian alcoholic drinks and mixer brands from 1996 to 2015. After a year of almost no advertising, average sales had fallen 16%, after two years 25%, after three 36%; larger growing brands often kept growing, while almost every smaller growing brand began to decline at once (Hartnett et al., 2021). This is yearly observational data, not an experiment, on broad-reach advertising across several media, and the summary I read does not single out search ads. But it shows that cutting to zero and trimming are different decisions.

## How do you check the worst case before cutting?

Set your loss limit first, then size the cut so that the worst case is bearable and reversible. The reason not to wait for a precise measurement: across Lewis and Rao's 25 experiments the median confidence interval on ROI was more than 100 percentage points wide (Lewis and Rao, 2015, abstract). The six steps below are mine.

1. Name the goal and the constraint (Goldratt's first step), then ask what the line feeds. If it feeds the constraint, for example new customers when you are short of customers, do not cut before you have separated that part out.
2. Split the spend by who receives it: new and light buyers against regular ones. That split is what produced the eBay result.
3. Compute an upper bound on the loss: the sales the platform credits to the spend, times gross margin. Blake and colleagues argue that conventional attribution exaggerates the effect (Blake et al., 2014, p. 3), so treat the number as a ceiling, not an estimate.
4. Cut a slice, reversibly, with a comparison group. eBay switched ads off in only about 30% of areas, to limit the scope of the test and the potential cost in case the ads proved considerably profitable (Blake et al., 2014, p. 13). By my reasoning, a keyword bid can be switched back on within a day while brand memory cannot, so cut harder-to-reverse channels more gently.
5. Fix the stop rule beforehand in terms of the loss you can bear, not statistical significance: for example, switch back on if the cut group trails the comparison group by more than a set threshold for two weeks running.
6. After each cut, return to step one. Goldratt warns against letting inertia turn old rules into the constraint (Goldratt, 2018, p. 4). Put the released money into the constraint, or label it plainly as a saving.

**Hypothetical example, continued.** The same shop sees its platform credit £30,000 of monthly revenue to the ads. At a 40% margin, the ceiling on lost gross profit is £12,000. Assuming the loss scales with the share cut (my assumption):

| Option | Monthly saving | Maximum gross profit lost | Worst outcome |
|:--|--:|--:|--:|
| Cut everything | £6,000 | £12,000 | -£6,000 |
| Cut one-third | £2,000 | £4,000 | -£2,000 |

Cutting everything could cost £6,000 in the first month, so it fails the check if your loss limit is below that. Cutting a third risks at most £2,000; if cash can stand it, that is a reasonable size, and on the earlier hypothetical table the result could equally be a £1,600 gain. The £30,000 is deliberately set above the £17,000 of the earlier table: here the platform credits more than the true effect.

## Going deeper: origins, critiques and limits

The theory of constraints began in manufacturing. The Goldratt text I read is adapted from his 1990 book *What Is This Thing Called the Theory of Constraints*, revised in 2018, and says the theory was first introduced in the business novel *The Goal* (Goldratt, 2018, p. 6). He asserts that every system has very few constraints and at least one, relying on the Boy Scout analogy in *The Goal* (Goldratt, 2018, p. 3); I have not read it. The constraint of an advertising budget is also not as clear-cut as a machine.

The eBay experiment covers one large brand, in the US, over the short term; the authors say the -63% is only a short-term estimate (Blake et al., 2014, p. 26). Nosko and Tadelis were employed by eBay Research Labs at the time (Blake et al., 2014, p. 1). The test areas were not purely random; a check using a randomly assigned candidate group gave smaller but noisier estimates (Blake et al., 2014, pp. 13, 31). The paper was later published in *Econometrica* 83(1), pages 155 to 174, according to Crossref; I have not read the published version, so every figure is from the working paper.

Lewis and Rao rest on 25 display-advertising experiments with large US retailers and financial firms, and say their results are unlikely to hold for firms with low sales volatility or for products nobody yet knows (Lewis and Rao, 2015, section 1). They discuss measurement, not how much to cut. Hanssens writes a commentary, not new research, and says response is "typically" concave (Hanssens, 2024). That a light cut on the rising part of an S could remove the effect is my inference, not his.

P&G and Chase carry a storytelling bias: I have only companies telling us the cut turned out fine, with no sample of cuts that went wrong. I also have no source on Vietnamese firms, and the post covers advertising spend only, not headcount or inventory.

What did I read? Blake and colleagues: the main text of NBER Working Paper 20171 (May 2014), with the appendix read in part; page numbers are the sequence numbers of pages in the NBER PDF file, cover included, and I have not read the *Econometrica* version. Lewis and Rao: the first half of the authors' draft of 18 September 2014, not the rest and not the *Quarterly Journal of Economics* version; section numbers follow the draft. Goldratt (2018): the first six pages of the PDF; the seventh would not extract. Hanssens: the full text of the 2024 copy and the 2023 draft. Bruell and Terlep (WSJ) and Maheshwari (NYT): in full, through the Fox Business and CNBC reprints, because the original pages block automated access. Kirkpatrick and Tode (Marketing Dive): in full. Hartnett and colleagues: only the summary on the ARF page; the DOI was checked through Crossref. I could not open the Ad Age article (the source of the $140 million figure) or the Reuters report that P&G cut $200 million over 2017, so I cite neither, and I do not cite Goldratt and Cox (1984) because I have not read it. All sources date from 2014 to 2021, and I have not checked whether P&G or Chase changed approach afterwards.

## Frequently asked questions

### Should I switch off ads on keywords that contain my own brand name?

For eBay, a brand everyone knew, switching them off lost almost no traffic in the short run (Blake et al., 2014, p. 10). The authors name one possible reason to keep them, which they call perhaps the only reasonable argument: rivals may bid on your brand terms to divert your visitors (Blake et al., 2014, p. 5). They report that once eBay stopped bidding, rivals were not seen doing this, though that is an observation rather than a measurement, and some retailers kept bidding on phrases such as ebay shoes (Blake et al., 2014, p. 24). If a competitor is buying your name, try a switch-off on a small slice first.

### Can a small firm test this when it is too small to run an experiment?

Lewis and Rao say that telling a campaign that wastes all its money from one that pays back typically takes experiments in the millions of person-weeks (Lewis and Rao, 2015, section 1). A small firm is almost certainly below that scale, so do not wait for statistical significance. That is my inference from their paper, which does not discuss small firms; what you can do is size the cut so the worst case stays bearable.

### If I do not know my firm's constraint, should I still cut?

Goldratt makes identifying the constraint step one of his five focusing steps, after the system's goal has been defined (Goldratt, 2018, pp. 2-3). Without it the cut may still be right on profit, but do not call it moving resources to the bottleneck. Call it a saving, and keep the cut small. That distinction is mine; the pages of Goldratt I read do not discuss advertising spend.

## References

- Blake, T., Nosko, C. and Tadelis, S. (2014) *Consumer heterogeneity and paid search effectiveness: a large scale field experiment*. Cambridge, MA: National Bureau of Economic Research (NBER Working Paper 20171). https://doi.org/10.3386/w20171
- Bruell, A. and Terlep, S. (2017) *P&G cuts more than $100 million in 'largely ineffective' digital ads*. Wall Street Journal (Dow Jones Newswires), 27 July 2017; syndicated copy on Fox Business. Available at: https://www.foxbusiness.com/features/pg-cuts-more-than-100-million-in-largely-ineffective-digital-ads (Accessed: 2026-11-03).
- Goldratt, E.M. (2018) *The 5 focusing steps of the theory of constraints (first published 1990 in What is this thing called the theory of constraints)*. North River Press (free download, adapted from the 1990 book). Available at: https://northriverpress.com/wp-content/uploads/2018/01/Free-download-5FS.pdf (Accessed: 2 November 2026).
- Hanssens, D.M. (2024) *Using return on marketing investment effectively*. Impact at JMR, Journal of Marketing Research (American Marketing Association). Available at: https://www.ama.org/marketing-news/using-return-on-marketing-investment-effectively/ (Accessed: 2026-11-03).
- Hartnett, N. et al. (2021) ‘When brands go dark: examining sales trends when brands stop broad-reach advertising for long periods’, *Journal of Advertising Research*, 61(3), pp. 247–259. https://doi.org/10.2501/JAR-2021-009
- Kirkpatrick, D. (2017) *P&G's sales rise despite trimming $140M from digital ad spending*. Marketing Dive, 28 July 2017. Available at: https://www.marketingdive.com/news/pgs-sales-rise-despite-trimming-140m-from-digital-ad-spending/448124/ (Accessed: 2026-11-03).
- Lewis, R.A. and Rao, J.M. (2015) ‘The unfavorable economics of measuring the returns to advertising’, *Quarterly Journal of Economics*, 130(4), pp. 1941–1973. https://doi.org/10.1093/qje/qjv023
- Maheshwari, S. (2017) *Chase had ads on 400,000 sites. Then on just 5,000. Same results.* The New York Times, 29 March 2017; syndicated copy on CNBC, 31 March 2017. Available at: https://www.cnbc.com/2017/03/31/chase-had-ads-on-400000-sites-then-on-just-5000-same-results.html (Accessed: 2026-11-03).
- Tode, C. (2017) *P&G's Marc Pritchard on where digital marketing is headed now that there's more transparency*. Marketing Dive, 20 November 2017 (interview). Available at: https://www.marketingdive.com/news/pgs-marc-pritchard-on-where-digital-marketing-is-headed-now-that-theres/511237/ (Accessed: 2026-11-03).

---

Canonical HTML version: https://staging.khoi.co.uk/en/blog/cat-ngan-sach-don-bay-bao-hoa