The next euro first pays for the two decisions a test can settle: paid search on your own brand, and your SEO's exposure to AI Overviews. After that comes a capped GEO budget, described as what it is today: a brand bet, with exit conditions written in advance.
The argument here is not that people have stopped using AI assistants. Usage is climbing fast. The argument is that the evidence is lopsided. Brand search has been tested by experiment since 2012. The effect of AI Overviews on clicks has a randomized experiment behind it. For GEO, no incrementality test on revenue could be identified as of 4 November 2026.
Funding these lines on the same basis confuses what has been measured with what is hoped for. The order below keeps them apart.
What is actually known about each budget line?
Very different amounts, and that gap should drive the allocation.
| Line | Best public evidence | Strength | Default decision |
|---|---|---|---|
| Paid search on your brand | Experiments at eBay (2012) and across thousands of brands on Bing (2018) | High, but dated | Test market by market |
| Non-brand paid search | Geographic experiment at eBay: new and infrequent customers respond, frequent buyers do not | High, but one very well-known brand | Measure by customer type |
| SEO | Randomized experiment: 39.8% fewer organic clicks where an AI Overview appears | High, but US, desktop, working paper | Protect, then redirect |
| Brand and PR | Correlation of 0.66 to 0.71 between web mentions and AI mentions across 75,000 brands (Ahrefs) | Low on causation | Keep, add a criterion |
| GEO | One quasi-experiment on a single domain, placebo not significant | Very low | Cap, fund in stages |
| ChatGPT ads | No public incrementality test identified | None | Optional, isolated test |
Reading based on the sources cited in this article, as of 4 November 2026.
None of this says GEO is worthless. It says GEO is the line with the least evidence behind it. Olivier Martinez's July 2026 review of 45 studies, a preprint, gives "very low" confidence to the link between citations and clicks, conversions or revenue. The best-documented quasi-experiment, by Watanabe and Nakayashiki (June 2026, one domain), estimates a 1.82x lift in ChatGPT referrals. Its placebo test returns p=0.16, and the authors call the effect suggestive rather than conclusive.
Does AI traffic justify a budget on its own?
No, not as of 4 November 2026. Across 118,671 sites tracked by Ahrefs in October 2026, AI assistants account for 0.46% of visits and search for 27.55%, about 60 times more. On a different panel, SE Ranking puts the EU at 0.27% for January to April 2026. For France, Kevin Indig (Growth Memo, October 2026) puts these visits at 2.44% of SEO traffic, using Similarweb estimates.
A higher conversion rate would not change the order of magnitude, and it is not a given anyway:
| Study | Sample | Result against organic |
|---|---|---|
| Kaiser and Schulze, Aug 2024 to Jul 2025 | 973 e-commerce sites | Organic converts roughly 13% better than ChatGPT |
| Amsive, Sep 2025 | 54 sites | AI 4.87%, organic 4.60%, not significant (p=0.794) |
| SEOWorks, Sep 2025 to Aug 2026 | 63 of 95 sites | 1.18x organic, 0.67x paid search, 0.62x organic for B2B product businesses (10 sites, median 0.84x) |
On a last-click view, the case is thin. Usage tells another story. The Reuters Institute (YouGov, August to September 2026) reports a 21-point rise in weekly use in France in a year.
That gap between usage and clicks leaves two explanations open. AI shapes decisions before any click happens, or it weighs less than its audience suggests. Kevin Indig looked for the link across four brands, running 632 lag tests between AI visibility, direct traffic and branded search (October 2026). He could not establish a reliable link, with few weekly observations, and notes an effect could exist that he failed to isolate. Rand Fishkin (SparkToro, October 2026) argues for treating AI visibility as a brand channel rather than a traffic channel. That is an opinion, and it is the reading used here, because it forces you to say what you are buying.
Should you cut brand search to fund GEO?
Not without a test, and not in GEO's name. Recommendation Visibility is how often your brand naturally appears when a user expresses intent to an AI. The idea that a rise in it lets you lower your brand bids is attractive, but no published study could be identified that has tested it. The question is worth asking, but no saving can go in the budget on that basis yet.
The value of brand search itself has been measured, and the results disagree in useful ways.
| Study | Setting | Result |
|---|---|---|
| Blake, Nosko and Tadelis, Econometrica, 2015 | eBay, brand bidding stopped on Yahoo and MSN in 2012 | 99.5% of traffic retained, picked up by organic |
| Simonov, Nosko and Rao, Marketing Science, 2018 | Thousands of brands, experiments on Bing | With no rival, a 1% to 4% click lift. With no ad against rivals, they capture 18% to 42% of clicks |
| Chan and Van Alstine, Google, 2012 | 390 pause studies, clicks only, brand queries not singled out | 50% of paid clicks incremental with organic in first place, 82% in positions 2 to 4 |
The larger the brand, the less the ad adds. A rival in the auction changes the decision: Simonov and colleagues then find defensive bidding strongly profitable. But all of this predates AI Overviews and automated campaigns, and the last study comes from the company selling the ads, so none of it transfers directly to your account. Google's own authors recommend randomized experiments, geographic ones for example.
Four rules frame the test:
- Split the queries. With and without a competitor in the auction: these are two separate decisions.
- Randomize the regions. Some with no brand bidding, others unchanged, for a period that spans your buying cycle.
- Measure the total. Overall brand traffic and sales. Paid clicks fall to zero by design and tell you nothing.
- Close the leaks. Check that your automated campaigns do not pick up the queries you removed.
If the test frees up budget, that money funds what comes next. If it shows the spend is useful, you finally have a number that justifies keeping it.
ChatGPT advertising calls for the same discipline. According to Metaturbine's analysis (September 2026), ads have been serving in France since 24 August 2026, with no keyword bidding and no way to exclude existing customers, and performance claims are unaudited. If you test this channel, keep it isolated.
Is SEO still the first line to protect?
Yes, though not at its current scope. Agarwal and Sen's experiment (1,065 US desktop users, January to February 2026, a working paper reported by PPC Land) hid AI Overviews at random. Where they appear, organic clicks drop 39.8%. The effect sits on informational queries. On navigational and transactional queries it cannot be told apart from zero, though those estimates are noisier. Sponsored clicks show no significant change.
France has only just entered this picture, with the launch on 22 July 2026. SE Ranking found an AI Overview on 52.63% of 100,000 French keywords in early August, from 87.1% in business to 20.28% in fashion and beauty. That is a share of keywords, not of searches. Across 963 large domains, Ahrefs reports a median decline of 8.8% in clicks and 5.7% in click-through rate, with click-through rate down 23.1% for domains where over 20% of queries trigger an AI Overview, on nine days of data. These first French figures still need confirming.
The opposite story is just as shaky. Kevin Indig estimates French organic visits fell 24.4% between August 2024 and August 2026, against 13.2% in the US, while AI Overviews arrived at the very end of that window. He notes the gap narrows substantially on identical sites and that his comparison cannot isolate their effect. Over the same period, by his estimate, paid gains offset only about 6.7% of French organic losses. Buying back what SEO loses is far from automatic.
The decision is to sort your queries by intent and exposure before moving any money. Generic informational content under heavy exposure yields less. Commercial pages and data only you publish keep their value either way. This work also serves GEO: SEOWorks finds a 0.534 correlation between AI referral growth and organic growth across 61 sites after controlling for growth in the rest of the traffic, and C-SEO Bench (NeurIPS 2025) finds conventional SEO, which improves a source's position, markedly more effective than rewriting. Cutting SEO to fund GEO would probably weaken both.
Do brand and PR deserve more money because of AI?
Not for that reason alone. Across 75,000 brands (Ahrefs, December 2025), AI mentions correlate at 0.66 to 0.71 with web mentions, 0.35 to 0.47 with branded search volume and 0.22 to 0.29 with ad traffic. The sample filters for established domains, and the authors warn that correlation does not mean raising these metrics lifts AI visibility. A large brand piles up all of them at once.
The useful decision is smaller: keep your plan and add one selection criterion, the outlets and comparison sites engines actually read on your topics.
How do you size a first budget without inventing a benchmark?
Start from your stakes, not a percentage. The number in circulation, 12% of digital budgets going to AEO in 2025, comes from a survey Conductor published in April 2026, covering 250+ digital leaders at US companies with 500+ employees. It tells you what respondents say they spend, not what pays back.
Three inputs are enough:
- Exposure. Since 31 August 2026, Search Console has shown every site its impressions in generative AI features. Clicks from AI Overviews stay blended into Web traffic, so this loss can be estimated but not read directly.
- The floor. The revenue GA4 assigns to AI traffic on last click. It may understate influence, but it is the only observed number.
- The bet. What you are willing to lose to be present in these answers, in case their brand effect exists. This cannot be calculated. It gets decided and written down.
| Alba & Co | Assumption | Annual revenue |
|---|---|---|
| Revenue entering through informational organic | From the account | €600,000 |
| Share exposed to AI Overviews | 30% | €180,000 |
| Loss if the 39.8% effect applied as is | Measured in the US, on desktop | €71,640 |
| Share recovered through presence in answers | One quarter, favorable case | €17,910 |
| Last-click AI revenue, doubled in a year | €30,000 today | +€30,000 |
| Total, then margin at 50% | €47,910, so €23,955 in margin |
Illustrative example, fictional brand. None of these amounts is a benchmark.
The recovered quarter is generous: Pew (900 US adults, March 2025) found people clicked a link inside the AI summary on 1% of visits to pages showing one. In this example, clicks alone justify about €24,000 a year, internal time included. Past that point, traffic no longer pays for it. Brand may, provided the plan says so: a brand bet, this amount, this stop condition.
Split the cap across measurement, one owner's time, and the content or sources to win, which often already sit in SEO and PR budgets. The first stage covers the 90 days. The second opens only after the tests are read.
Who should own GEO in the organization?
One person, usually the head of SEO, with the CMO as arbiter. This rests on reasoning: no reliable survey could be identified.
SEO, because AI gains track organic gains in the available data, and the pages engines read are often pages you can win. Not SEO alone, because much of what engines read belongs to someone else: in the University of Toronto study (September 2025, API access, US queries), in the answers of a web-enabled GPT, third-party editorial sources make up 81.9% of citations in automotive and 72.7% in software, against 45.1% and 45.4% on Google.
That gives a working rule: the center of gravity follows the source map. If rivals and your own site dominate what gets read on your topics, SEO leads. If press and comparison sites dominate, PR leads the fieldwork. If communities and reviews dominate, brand and customer service join the loop.
| Topic | Owner | Contributes |
|---|---|---|
| Budget, stages, stopping | CMO | Finance |
| Prompts, reading, action plan | Head of SEO | Product marketing |
| Third-party sources | PR and brand | SEO |
| Brand search test, ChatGPT ads | Head of paid | Data |
| Test design and reading | Data | Finance |
| Accuracy of what AI says about you | Product marketing | Legal, customer service |
What matters most: the team doing the work does not get to grade it alone.
What should you stop doing?
Five habits to drop this quarter.
- Justifying GEO with a conversion multiple. The three studies above are enough to rule it out as a budget assumption.
- Funding llms.txt, special markup, chunking or rewriting "for AI". Google's guide, read on 4 November 2026, says none of this is needed for its search, and treats a page for every query variation, when done to manipulate, as a violation of its scaled content abuse policy. C-SEO Bench finds most rewriting methods ineffective and often harmful to ranking. See the tactics ranked by level of evidence.
- Showing an AI rank to the board. Across 117 queries repeated three times, Uberti-Bona Marin and colleagues (September 2026) measure recommended-product overlap of 0.178 on ChatGPT and 0.421 on AI Overviews, while naming at least one brand flips on only 11% to 13% of queries. Presence holds up better than order.
- Blaming AI for the whole organic decline. French data does not allow it.
- Defending or cutting brand bidding on conviction. Either way, the test costs less than a mistake renewed every year.
What does a 90-day plan look like?
Days 1 to 30: set up measurement and write the rules.
- Name the owner and write decision thresholds before seeing any data.
- List around 100 decision questions without your name, across at least 10 topics, per market, with competitors tracked, and start measuring on day 1: this is the GEO test's baseline.
- Isolate AI traffic in GA4, pull generative impressions from Search Console, and add "How did you hear about the brand?" to forms, knowing those answers stay noisy.
- Sort brand queries by competitor presence.
Days 31 to 60: launch two tests.
- The brand search test, under the four rules above.
- A GEO test with a control group, with treatment starting around day 57 after eight baseline weeks. You cannot switch off presence in an AI answer by region, so the control is built by topic, acting on only a randomly drawn half of a comparable set. Watanabe and Nakayashiki show why: their untreated pages also grew 3.5x, against 5.7x for the domain as a whole. Without a control, you credit your work with the platform's growth. The method is in proving GEO incrementality.
Days 61 to 90: read and decide.
- Read rates over several weeks. This article's arithmetic, at 95% confidence and 80% power: telling 30% from 40% takes about 350 independent observations per period, and answers to the same prompt are not independent. A two-point target makes no sense. See how many prompts a reliable measure takes.
- Annotate breaks. A model update or a measurement artifact can move visibility with no change on the brand's side: see false signals in AI visibility.
- Pick an exit written down on day 1. Raise the cap if the treated group pulls away from the control and a demand signal agrees. Hold if presence grows with no downstream signal, and own the brand bet. Fall back to measurement only if nothing separates from the control after a second cycle. This interim read decides the next tier. The test itself concludes only at its final read, after twelve treatment weeks.
At 90 days, the brand test can reach a conclusion if volume is sufficient. The GEO test, about five treatment weeks into twelve, gives a direction and no revenue effect. Tell the board on day one.
Which gaps belong in your decision memo?
Three, as of 4 November 2026. GEO's effect on revenue and the link between AI visibility and brand search: no public test identified. The French case: nine days of measured clicks in the most direct study, no France-specific B2B study identified, and a causal experiment run in the US on desktop, without mobile apps.
The thankless part of this plan is reading these signals side by side, for months. Brand Radar brings together both terms of Brand Demand Efficiency, the ratio between what you spend on your name and the real organic demand behind it, from Google Ads and Search Console, and shows who bids on your brand week after week. The decision stays yours, and Nanga never changes your accounts. Nanga's GEO module follows your prompts run after run, weekly by default and daily as an option, and measures your AI Visibility Index engine by engine, against your competitors. Its AI Traffic page reads in GA4, once the connection is set up with the Nanga team, the visits assistants send you.
Spark's SEO, GEO and acquisition experts, from the agency that builds Nanga, can frame these tests and carry out the work with your team.
To see your brand spend against your real demand, and your presence in AI answers: request a demo.
Frequently asked questions
What share of the marketing budget should go to GEO?
No reliable percentage exists. The 12% of digital budgets Conductor reported in April 2026 is self-declared spending from a survey of large US companies. Start from your AI Overview exposure and measured AI traffic, then set a cap and funding stages.
Does GEO have a proven return on investment?
No, as of 4 November 2026. Olivier Martinez's July 2026 review of 45 studies gives very low confidence to the link between citations and clicks, conversions or revenue. That does not prove there is no effect: no public test can measure it yet.
Can you cut brand bidding when AI visibility rises?
Nothing supports that claim: no published study could be identified that tests the link. Simonov, Nosko and Rao measured a 1% to 4% click lift with no competitor in the auction, and profitable defense when rivals bid. Test by region on your own queries first.
Who should own GEO in a marketing team?
One person, usually the head of SEO, with the CMO arbitrating budget. PR and brand take third-party sources, and the data team validates tests. If the press dominates what engines read in your category, the center of gravity moves toward PR.
What can you know after 90 days?
A brand search test can reach a conclusion if volume is sufficient. A GEO test with a control group gives a direction on presence in answers, not an effect on revenue. Its full protocol, eight baseline weeks and twelve treatment weeks, runs past the quarter.
Sources:
- Ahrefs, AI vs Search Traffic, October 2026 data.
- SE Ranking, AI traffic research study, June 2026.
- Kevin Indig, Growth Memo, Organic traffic is down, but it's not all AI, October 2026.
- Kevin Indig, Growth Memo, Does AI Drive More Direct Traffic?, October 2026.
- Rand Fishkin, SparkToro, Is AI Visibility a Marketing Boondoggle?, October 2026.
- Olivier Martinez, Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026), preprint, July 2026.
- Watanabe and Nakayashiki, Disentangling Answer Engine Optimization from Platform Growth, preprint, June 2026.
- Kaiser and Schulze, as presented by Frankfurt School, Where ChatGPT already works in online shopping, June 2026, and by Digiday, E-commerce sites see low sales from ChatGPT traffic, October 2025.
- Amsive, LLM traffic conversion rate study, September 2025.
- SEOWorks, AI Referral Traffic Benchmark, data from September 2025 to August 2026.
- Reuters Institute, Generative AI and News Report 2026, October 2026.
- Blake, Nosko and Tadelis, Consumer Heterogeneity and Paid Search Effectiveness: A Large Scale Field Experiment, NBER 2014, Econometrica 2015.
- Simonov, Nosko and Rao, Competition and Crowd-Out for Brand Keywords in Sponsored Search, Marketing Science, 2018.
- Chan and Van Alstine, Google Research, Impact of Organic Ranking on Ad Click Incrementality, March 2012.
- Agarwal and Sen, SSRN working paper, reported by PPC Land, Researchers find Google AI Overviews cut publisher clicks 39.8%, 2026.
- Metaturbine, ChatGPT Ads launched in Europe: here is what actually shipped, September 2026.
- Google France, announcement of AI Overviews and AI Mode in France, July 2026.
- SE Ranking, AI Overviews France study, August 2026.
- Ahrefs, 1 mois d'AI Overviews en France et déjà -23,1% de taux de clics, August 2026.
- Pew Research Center, Google users are less likely to click on links when an AI summary appears in the results, July 2025.
- Ahrefs, AI brand visibility correlations, December 2025.
- Chen, Wang, Chen and Koudas, University of Toronto, Generative Engine Optimization: How to Dominate AI Search, preprint, September 2025.
- Puerto et al., C-SEO Bench: Does Conversational SEO Work?, NeurIPS Datasets and Benchmarks, 2025.
- Uberti-Bona Marin et al., Auditing AI-Generated Product Recommendations, preprint, September 2026.
- Conductor, The State of AEO / GEO in 2026: CMO Investment Report, April 2026.
- Google Search Central, generative AI reports in Search Console, June 2026, AI features and your website and AI optimization guide, read November 2026.