Why guarantees in SEO should worry you

This is not a claim that everyone offering one is dishonest. Some sell it because buyers ask for it, and buyers ask for it because they have been burned and want their risk covered. That instinct is correct. The instrument is wrong.

A guarantee feels like protection. Read closely, it is usually a transfer of keyword selection from you to the vendor.

What is actually being promised?

Find the definition and read it twice. It is always narrower than the headline.

Check which search terms it covers, and who picks them. Check the geography, since a position in one part of town is not the position across the city. Check whether the map results count toward the promise, because those follow different rules from the blue links. Check whether it says top ten, top twenty, or first page, and on which device.

The common shape is a promise about terms that carry your business name, or terms so specific that almost nobody types them. Both are winnable in weeks. Neither produces customers.

A promise you can satisfy with a term nobody searches is not a promise, it is a definition. Ask for the exact list of terms in advance, with monthly search volume next to each, and watch what happens to the offer.

Why can nobody promise a ranking?

Because the result is not one thing anymore, and one of the largest inputs is something neither of you controls.

In Whitespark's 2026 Local Search Ranking Factors survey, 47 local search practitioners scored 187 factors. Proximity of Address to the Point of Search came second, at 225 out of a possible combined score, behind only the primary business category at 227. The second most important input to your map position is where the customer happens to be standing. Two people a mile apart get different results for the same words. A screenshot from the agency's office proves very little about what your customers see.

Grade that survey honestly. It is expert opinion, not test data, and its author says so. But this particular finding is not controversial and you can verify it yourself in twenty minutes by searching the same term from two ends of your service area.

The generative answers on top have made this looser again. Researchers at the University of St. Gallen ran identical prompts across four AI engines daily for 45 days and then repeated identical prompts within a single day. Between two consecutive days, only 34% to 42% of cited sources overlapped. Within a single day, repeated runs of the same prompt shared 32% to 43% of sources, and on ChatGPT the source overlap measure came in lowest of the four engines tested. ChatGPT activated web search on only 42.2% of runs, returning no citations at all on the rest.

Underneath that is a property of the models themselves. A variance decomposition of 12,933 brand responses attributed 34.8% of total variance to repeating the same prompt alone, and one lab ran an identical prompt a thousand times at temperature zero and got 80 distinct completions.

So a promise about appearing in an AI answer is a promise about a system that does not repeat itself. Even the inputs that plainly do influence those answers behave this way, and reviews feed AI answers without any mechanism to promise inclusion.

There is one exception worth naming. Ad platforms sell placement directly, and cost-per-lead products exist where you pay only when a lead arrives. Those are real structures with real accountability. They are not organic ranking, and they should not be described as if they were.

How many positions are even left to promise?

Here is the arithmetic, and it is the part of this conversation nobody has with buyers.

Count the slots. A classic local pack has three results. Sterling Sky reports that AI-powered local packs, appearing on roughly 7% of tracked keywords, show one or two businesses instead of three and carry no call buttons. On those queries, the number of available organic positions has fallen by at least a third before anyone optimizes anything.

Count the businesses. A Places Scout analysis found AI local packs surfacing 5,943 unique businesses against 18,330 in regular three-packs, roughly 32% as many. Across 322 markets, 88% showed fewer unique businesses in the AI pack than in the standard one.

Count what got added above you. Sterling Sky's tracking shows local pack ads going from about 1% of mobile reports in early 2025 to about 22% by December 2025, and Local Services Ads from about 11% to 31% of tracked queries.

Now price the promise. Even if the agency delivers position one, nobody can tell you what it is worth. Google publishes no local pack position dimension anywhere. Its Business Profile performance reports give views, calls, website clicks, direction requests, messages and bookings, and no rank at all. Search Console has no local pack dimension either. The only traceable click-share study of that shape is a 2018 lab test on five San Francisco desktop results pages, and three of the four figure sets in circulation trace to studies that were never published.

So the honest sentence is that the position is promised, the number of positions is shrinking, and the value of the position is unpublished. A clause can be satisfied inside all three of those facts without your calls changing at all.

That is not an argument for pessimism. Position one still takes materially more traffic than position three, and the work that gets you there is real. It is an argument against writing a contract around a number that carries this much unmeasured space inside it.

What does a guarantee do to the work itself?

It changes what gets worked on, quietly and immediately.

Once the fee depends on hitting a defined position, the rational move for the agency is to pick the terms most likely to hit and spend the hours there. That is not villainy, it is incentive design. You get effort aimed at the contract rather than at your revenue.

The AI version of this deserves its own warning, because the measurement is soft enough to be gamed without anyone intending to. If a clause pays out on a visibility score, ask two questions before you sign: how many runs per prompt per day, and over what window. The St. Gallen paper's own prescription is at least seven runs per prompt per day, reported on a two to four week rolling aggregate, never week over week, because the standard error of a detection rate is around 0.10 at seven runs. At that precision, a score moving from 40% to 45% in a week is statistically indistinguishable from no change. A clause built on a weekly reading of a single prompt is measurable only in the sense that a number will exist.

Google's own documentation is worth having on hand for the tooling conversation: be wary of third-party tools that promise ranking success or claim to use internal Google metrics, because no third-party tool has access to its internal ranking or AI systems.

The version I would want instead is boring. Agree on the two or three searches that actually bring you paying work, agree what happens by when, and accept that positions on those terms may take a while. That is a harder conversation and a better engagement.

How the fee is built has more effect on behavior than any promise attached to it, and the common pricing structures each push work in a different direction.

Is a money-back offer any better?

Somewhat, and less than it sounds.

A refund returns the fee. It does not return the quarter, the staff time you spent in meetings, the content that now sits on your site in someone else's voice, or the position you did not gain while a competitor did. For most small businesses the fee is the smallest of those losses.

Read the trigger conditions too. Refund clauses usually require you to have supplied everything on time, approved everything within a stated window, and made no changes to the site. Those conditions are reasonable on their face and they also make the clause difficult to invoke.

And check what happens to the assets. A refund with no ownership clause can leave you paid back and locked out, which is why who owns the site, the content, and the accounts is the term that matters most when things end.

What can an agency honestly promise?

Quite a lot, and all of it is checkable.

They can commit to deliverables with units. Twelve listings corrected. Four pages published. Your profile's services and products built out properly, which is real work with a visible result, and a services list done correctly is something you can inspect yourself the same day.

They can commit to response times, to a reporting date every month, to who does the work, to sending you the analytics access, and to a named baseline recorded before anything changes.

They can commit to what is genuinely measurable in AI search, which is a shorter list than most proposals imply and is not nothing. Crawler hits from AI bots in your own server logs are deterministic. Google Search Console publishes generative AI impressions by page, country, device and date, though it publishes no clicks, no click-through rate, no position, and not the prompt anyone typed. A citation rate for a defined prompt set is reportable as a probability with a confidence interval if the run count and window are stated. Anything beyond that is not measurement.

They can commit to a review point. Not a result, a review. At ninety days, here is what we said would be true, here is what is true, here is what we do next. What ninety days can and cannot produce is knowable in advance, and agreeing to it in writing gives you more protection than any promise clause.

What to do this week

If you have an offer with a promise in it, ask three questions in one email.

Which exact search terms does this cover, and what is the monthly search volume for each. What is the geographic radius and the device the check is run on. What do I have to have done for the clause to remain valid.

Then run the proximity test yourself. Search your main term from your own address and from the far edge of your service area. If the results differ, and they will, you have just demonstrated why the clause needs a radius in it.

Then ask the question that matters more. "If we hit those positions and my calls do not increase, what changes?" A good answer describes a conversation and a plan. A weak answer explains that the contract was satisfied.

Questions about a guarantee clause you were sent?

Email me at eric@seod.com with the paragraph pasted in, and I will tell you what it actually obligates the agency to do and what you would have to prove to invoke it. Plain language, no commentary on whether you should hire them. I am not asking for the business and I will not send anything after the reply.

There is more on hiring and working with agencies alongside this.