Proving the value of SEO means separating what the work on the site produced from what would have happened anyway: the season, an ad campaign launched the same month, a competitor closing down, a Google update. A rising chart isn't proof. It becomes proof when you compare it with something nobody touched.
On 10 October 2026 we read the first page of Google UK for "how to prove the value of seo". There's a list of five arguments for SEO (seoClarity), five steps for measuring it (Precis), seven ways to show it, from KPIs to ROI and attribution (HawkSEM), a piece on proving value when rankings aren't enough, and Google's SEO Starter Guide. HawkSEM admits that algorithm changes, competitors and shifts in behaviour make the effect of SEO hard to isolate, but none of them explains how to isolate it, which is the first thing a finance director asks. The closest result to the right question is a Quora thread that simply asks, "How do you prove the value of SEO?"
This article starts there. Before the contract you need a business case, and we wrote it up in how to explain SEO to your boss. Every month you need an SEO report. Here we deal with the third document, the one that comes after six or twelve months and has to stand up to a single question: how do we know it was SEO?
Why a rising chart proves nothing
Three things move organic traffic without SEO having anything to do with it. Anyone who has read a few reports knows them.
- The season. A site selling air conditioning grows in June even if nobody works on it. Comparing July with March measures the heat, not the optimisation.
- Advertising. A campaign on Google Ads, on social media or on TV increases searches for the company name, and those searches end up in organic traffic. The client sees more visits and more orders and gives the credit to whoever does their SEO.
- Google. An algorithm update, a competitor hit by a penalty, a results page that changes shape all move clicks one way or the other.
Since 2025 a fourth has been added, and it works the other way: generated answers. The Pew Research Center recorded 68,879 searches made by 900 US adults in March 2025. With an AI summary on the page, a click on a traditional result happened in 8% of visits, against 15% for searches without a summary (Pew Research Center, July 2025). It's a US sample. AI Overviews reached the UK on 15 August 2024, and we've lined up the UK data available in the article on the impact of AI Overviews on SEO.
The consequence is that impressions and clicks no longer move together. On our own site, filtering Search Console to Italy only, our main market, impressions went from 1,799 to 4,668 (+159%) between September 2025 and July 2026, while CTR fell from 0.67% to 0.39%. The same report can tell a story of success or failure depending on which line you put at the top. That's why a serious goal carries two numbers, as we explain in the guide to SEO goals, and a serious proof carries a comparison.
The ladder of proof: what each one rules out
Not all proof of SEO value carries the same weight. We've ordered it from weakest to strongest. The column that matters is the third: which alternative explanation the proof manages to rule out.
| Proof | How it's done | What it rules out | What it doesn't rule out | When you can do it |
|---|---|---|---|---|
| 1. Rankings and tool scores | Screenshot of the rank tracker or audit tool | Nothing | Everything: a ranking isn't a customer | Always |
| 2. Before and after on the total | Month on month in Search Console or GA4 | Nothing systematic | Season, advertising, Google updates | Always |
| 3. Year on year | Same period of the previous year | Recurring seasonality | Advertising, updates, AI Overviews | With at least 12 months of data |
| 4. Share of visibility (share of voice) | Your visibility divided by that of the whole keyword cluster | Season and market trends, because they hit everyone | The quality of the traffic | With a rank tracker on a group of keywords |
| 5. Control group | Changed pages against similar pages left as they were | Season, advertising and updates that affect both groups | Differences between two badly chosen groups | With two groups of pages of the same type |
| 6. Counterfactual time series | A model estimates the traffic you would have had without the change (CausalImpact) | Trend and seasonality, if the control series are good | Events that affect only your site | With at least a year of daily data |
| 7. SEO A/B test | Pages on the same template split at random into two groups | Almost everything | Little, if the sample is large | With many pages on the same template |
| 8. Pausing ads | Campaigns paused on a group of queries or in some regions | The overlap between ads and organic results | Long-term effects | If you already run Google Ads |
Row 4 is the most underrated. The season hits every site in a sector at the same moment: if you grow in June and so do your competitors, your share stays flat and the rising chart says nothing about you. If your share grows, you've grown more than the market. It's the figure we use for Macropix, a Milan-based LED wall manufacturer and Visilay client: 25.55% share of voice on the LED wall keyword cluster against 18.08% for amazon.it (Semrush, July 2026), and the keyword "monitor pubblicitario" (advertising monitor) up from position 88 in 2020 to 2 in 2025. How to calculate the share is explained in the guide to share of voice.
Rows 5 to 8 are the four proofs that hold up even in front of people who check numbers for a living. Let's go through them one by one.
Proof 1: the control group
A control group is a set of pages similar to the ones you're working on, which you leave unchanged over the same period. What happens to them is the estimate of what would have happened to your pages without the work. The difference between the two growth rates is the effect of SEO. In statistics the method is called difference in differences, and it's the cheapest proof: all you need is Search Console and a spreadsheet.
An arithmetic example, to redo with your own numbers. An ecommerce site rewrites the copy on 40 categories and leaves 40 similar ones as they were. In the following three months, compared with the same quarter the year before, organic clicks look like this:
| Group | Clicks before | Clicks after | Change |
|---|---|---|---|
| Rewritten categories (40) | 1,000 | 1,300 | +30% |
| Untouched categories (40) | 1,000 | 1,150 | +15% |
| Estimated effect of the rewrite | about 150 clicks | +15 points |
The traditional report would have said +30%. Half of that growth would have come anyway, from the season or the market. Claiming all of it for SEO works until someone does the sum, and on the day they do, they stop believing the rest too.
Three rules decide whether the comparison holds:
- The control group is chosen before you make any changes, and written down somewhere with the date. Choosing it afterwards, looking at the data, means choosing the one that proves you right.
- The pages have to be of the same type and with similar traffic: product pages against product pages, articles against articles. The blog can't act as a control for the categories.
- The control isn't touched. No new internal links to those pages and no changes to the templates they sit on, for the whole length of the comparison.
For a single URL the right source is Search Console, with the 28-day comparison and a verdict that arrives after four to eight weeks, as we explain in the guide to SEO analytics. The extra clicks then need translating into pounds with the value-per-click calculation you'll find in is SEO worth it.
Proof 2: the counterfactual time series
If the change affects the whole site (a migration, a new template, work on speed), there are no pages to leave out. In that case the point of comparison is built with a model. The most widely used is CausalImpact, described by Kay Brodersen and four other Google researchers in the Annals of Applied Statistics in 2015 (vol. 9, no. 1, pp. 247-274) and released as an open source R package.
Put simply: the model learns how your traffic moved before the change together with other series the change can't influence, for example interest in the category on Google Trends or the organic traffic of another country. Then it projects forward the curve you would have had without the change. The gap between the real curve and the estimated one is the effect, with an uncertainty interval around it. In the paper the worked example is in fact an online ad campaign and its effect on traffic from search.
The weak point is the control series. If you pick one the change has influenced, the model underestimates the effect. If you pick one that has nothing to do with your traffic, the model estimates nothing. Always report the interval: a hypothetical +12% with an interval running from -3% to +27% means "we don't know yet", and that's as useful a piece of information as a success.
Proof 3: the SEO A/B test
An SEO A/B test splits pages on the same template at random into two groups, changes something in only one and compares organic traffic. Unlike a classic A/B test you don't split users, because there's only one Googlebot: you split pages. You need a site with many pages that share the same structure, such as product pages, property listings or location pages.
The figure that changes how you present these tests comes from SearchPilot, the London company that specialises in them. Of the tests run in 2022 and 2023, about 15% came out positive with statistical significance, 7-8% negative and almost 75% inconclusive (Sam Nemzer, SearchPilot, April 2024). Craig Bradford, also at SearchPilot, puts it this way: "Most of your experiments are going to be negative or inconclusive".
Read from the value side, that 7-8% counts as much as the 15%. Every negative test is a change that without a test would have gone live across the whole site and taken traffic away without anyone understanding why. A testing programme proves value through the losses it avoids too, provided they go into the report with the same prominence as the wins.
Google allows these tests and sets the rules in the Search Central guidance on A/B testing: no cloaking, meaning Googlebot and users see the same versions; canonical tags on variants that have a different URL; 302 redirects, not 301, if the test redirects; and a test that lasts no longer than necessary. Google writes that an experiment left running too long may be interpreted as an attempt to deceive search engines.
Proof 4: switching the ads off
The most direct way to measure what an organic result is worth is to see what happens when the ad above it is gone. eBay did it. Tom Blake, Chris Nosko and Steven Tadelis paused ads on brand searches and measured that 99.5% of the clicks lost from the ads were immediately recovered by the organic result. eBay's annual paid search spend in the US was estimated at $51 million. The study came out as an NBER working paper in 2014 and later in Econometrica.
Google has published numbers that point the other way, and it's worth reading them together. Across more than 400 studies in which advertisers had paused their campaigns, 89% of ad clicks were incremental, meaning they weren't recovered by organic results (Chan et al., Google, 2011). In the follow-up study, on 390 pauses, the share fell to about 50% when the advertiser also ranked first organically, and stayed at 100% when there was no organic result of theirs. 81% of ad impressions appeared with no organic result from the same site (Chan et al., Google, 2012).
The two pieces of research don't contradict each other. eBay measured searches for its own brand, where the site is already first; Google measured the average across all queries. Together they say something that translates into money: where you rank first organically, part of the clicks you pay for you would have had anyway. That part is SEO value, and you prove it with a controlled pause of campaigns on a group of queries, or with the geo experiments Google described in 2011, switching ads on and off in different regions. These studies are more than ten years old: today's results page has more ads and generated answers. The method still holds; the numbers need measuring again on your own account. How the two channels share the work we cover in SEO and Google Ads together.
The reverse calculation is replacement cost: how much it would cost to buy through ads the enquiries that now come from organic search. For Macropix, between July 2025 and June 2026, Google Ads brought 374 conversions on €16,805 of spend, €44.93 per conversion. In management accounting terms, every qualified organic enquiry on that site is worth what it would cost to buy it back. If you want to take the same measurement on your own account, it's one of the checks we run in our Google Ads campaign management.
How to present proof of SEO value
Inside a business, the proof is read by someone who doesn't have time to redo the sums. According to the UK Business Data Survey 2026 (DSIT, fieldwork by Ipsos from October 2025 to January 2026, published June 2026), 86% of UK businesses handle digitised data, but only 25% of those analyse it to generate new insights. Among large businesses the share is 69%, among medium-sized ones 49%, small 35% and micro 26%. In a small business, whoever reads your proof often has no analyst next to them. That's why the proof fits on one page, with six rows that are always the same.
| Row | What it contains | Example |
|---|---|---|
| Hypothesis | What we expected and why, written down beforehand | Rewriting category copy increases non-brand clicks |
| Change | What changed, where, with the date | 40 categories, copy published on 3 March |
| Control | The point of comparison, chosen beforehand | 40 similar categories left untouched, chosen on 1 March |
| Result | The difference, with the interval if there is one | +15 points against the control, about 150 clicks in three months |
| Value | The result translated into money | Clicks times the value of a click, or enquiries times replacement cost |
| What we don't know | The limits, written down before anyone asks | Effect of AI Overviews on the group's queries |
The last row gives credibility to the other five. It's the same section we recommend in the monthly report, and in a proof it weighs more, because the reader is deciding whether to renew a budget. The value row needs careful handling too: the model used to attribute conversions can move it a lot, as we measured on three GA4 properties in the article on attribution models. For the actual return calculation, the formula is in our guide to SEO ROI.
When the proof can't be done
On a site with twenty pages and a hundred clicks a month none of the strong methods works: there aren't enough pages for a control group, nor enough data for a model. In that case it's better to say so and fall back on three things that stay honest.
- The year-on-year comparison, with seasonality stated and the period's ad campaigns listed alongside.
- Share of visibility on the keyword cluster that matters, compared with two or three named competitors.
- Enquiries in the CRM with the question "how did you hear about us?" on the form or in the first phone call. It's messy data, but GA4 can't give it to you: someone who searched on Google, found the company and then phoned doesn't appear in any Analytics report.
In B2B it's the most common situation. Of the nine keywords that bring Macropix the most qualified enquiries, none contains the company name: they're product searches such as "ledwall" (LED wall, 4,400 searches a month in Italy) or "totem led" (LED totem). Information like that is worth more than a traffic curve, because it says that without SEO those customers wouldn't have searched for the company by name. The full case is in the article on SEO for manufacturing.
Something the industry rarely says: the most convincing proof is the one that could have proved its author wrong. If SearchPilot sees three tests out of four end without a clear result, a consultant who only ever reports successes isn't showing they're better, they're showing they didn't measure. Proposing a control group first, before the client asks, costs a few wins in the reports and buys the only thing that gets a contract renewed after the first year: trust in the numbers. It's how we set up SEO consultancy at Visilay.
Frequently asked questions
It depends on the proof. On a group of pages the verdict comes after four to eight weeks of data, a year-on-year comparison needs twelve months of history, and enquiries and sales move after rankings. It pays to agree at the start which proof will be used and when.
No. A ranking proves the work was done, not that it brought in customers. With AI summaries the same ranking can bring fewer clicks than a year ago: in the Pew sample, clicks on a traditional result fall from 15% to 8% of visits when a summary appears.
Partly. GA4 shows the conversions attributed to organic search, but according to the attribution model chosen and without the phone calls and enquiries that don't go through the site. It needs cross-checking against the CRM and Search Console.
No. A control group comparison can be done with Search Console and a spreadsheet, and CausalImpact is open source, with no licence fee. Dedicated platforms such as SearchPilot are for sites with many pages on the same template that want to test continuously.
It depends on the queries. Pew found the AI summary on 60% of searches that start with a question and on only 8% of one- or two-word searches. Short searches, like a product name, are affected much less: one more reason to measure value in enquiries, not visits.
