What to do when you have been burned before

The instinct after a bad engagement is to look for a more trustworthy person. That is the wrong variable. Trustworthiness is real but unobservable in advance, and you already believed the last one.

What you can control is structure. A short first phase, a written starting point, assets in your name, and one artifact a month you can check yourself. Those work regardless of who you hire.

What do I do first, before hiring anyone?

Get your property back, and do it before you are in a dispute.

Domain registrar account. Hosting. Website admin. Google Business Profile owner role. Analytics and Search Console. Ad accounts. Any call tracking numbers printed on a vehicle, a sign, or a menu.

Then get copies. Content files, images, any research they produced, and an export of form submissions and customer data if the site collected any.

Do this while the relationship is merely disappointing rather than hostile. Requests get answered faster before anyone has said anything final, and ending an engagement cleanly is mostly a matter of sequence.

Recovering access is not confrontation. It is the same as taking your keys back from a contractor who finished the job.

How do I tell what actually went wrong?

Four failure modes cover almost all of it, and they need different fixes.

The scope was wrong. They did what was sold, and what was sold was not what your business needed. Common when a package is bought rather than a diagnosis. The fix is discovery before scope.

The work did not happen. Sold, invoiced, not produced. The fix is countable deliverables and a dated change log.

The measurement never existed. Something may have improved and nobody can tell, because no starting values were recorded. We had our own version of this. A case study of ours said applications had tripled, and nobody had written down the starting number, so it came down. The change was probably real. The claim was not defensible, and the failure was not that anyone lied. It was that nobody recorded day zero.

Expectations were wrong. The work was competent, it needed longer than you were told, and you stopped at month four believing you had been cheated. This is the most common of the four and the least often diagnosed, which is why it is worth reading what a first ninety days can honestly produce before you decide what happened to you.

There is a fifth possibility nobody enjoys. Approvals sat unanswered, photos never arrived, and production stalled on your side. If that is even partly true, name it now. You cannot fix a vendor problem that was actually a bandwidth problem.

Which protections actually prevent a repeat?

Five clauses and habits, all cheap to arrange at the start.

Ownership assigned to you on payment, covering the site, content, images, and configuration.

Owner-level access held by you throughout, not handed over at the end.

A short first phase with a defined finish. Ninety days is a reasonable length. You are buying a test of the working relationship, not only the work.

A baseline written down before anything changes. This costs an hour and it makes every future argument resolvable.

One monthly artifact you can check without help. Not a call, a document. Knowing what a monthly local search report should actually contain means you can specify it up front rather than accepting whatever template arrives.

What does a baseline actually look like?

Here is the hour, and then the arithmetic it makes possible.

Day zero. Write down six numbers. Calls in the last ninety days, from the phone system rather than from memory. Form submissions in the last ninety days. Bookings, jobs, or appointments closed. Views and actions from ninety days of Google Business Profile performance data. Your position on three named search terms, checked from three different points in your service area rather than from your own desk. Review count, plus the date of your newest review.

Twelve months later you have two columns. Say calls went from 210 to 256 across the same ninety day window. That is 46 more calls, roughly 22%.

Now apply the discipline that makes the number mean something. Add a third column: the same ninety days one year earlier. Was there a season in it? Did you change your hours, close for a remodel, or watch a competitor close? A 22% change with no seasonal control is not evidence, and the control column costs nothing because the data is already sitting in your account.

Then set the reading rhythm before you start, not after the first bad month. The measurement research in AI search is blunt about this and the logic carries over: report on a two to four week rolling aggregate, never week over week, because at seven runs per prompt the standard error of a detection rate is around 0.10, which means a five point weekly move is inside the noise. Local call counts behave the same way at small volumes. Ninety days is a signal. A week is weather.

One thing not to do with the baseline. Do not let anyone compare it to a published industry average. There is no current one worth using for local profile performance. The only large-sample study of Google Business Profile behaviour is from July 2019, on data from 2017 and 2018, built partly on metrics Google has since retired and on a definition of views that has since changed. Benchmark yourself against your own prior period and against your named local competitors. That is the whole method, and it is more defensible than any number a vendor can hand you.

What is the overcorrection to avoid?

Turning yourself into an auditor.

The natural response to being burned is to demand weekly proof. The effect is that your vendor starts optimizing for visible activity. You get screenshots, small changes, and a lot of reporting, and you get less of the slow work that actually moves anything. You will have made the relationship safer and the results worse.

Monthly is the right rhythm for evidence. Weekly is the right rhythm for questions.

The second overcorrection is buying on price. Getting burned once makes the cheap option feel like a limited downside. It is usually the same failure at a lower monthly cost, and now the work is thinner too.

The third is defensive decisions made out of suspicion rather than analysis. Owners who have been sold something they did not understand sometimes start refusing everything technical, including sensible things. Blocking AI crawlers is the clearest current example, and the evidence says it mostly does not do what people think it does.

BuzzStream analysed 4 million citations across 3,600 prompts covering ChatGPT, Gemini, Google AI Overviews and Google AI Mode. Among sites blocking ChatGPT's live retrieval bot, 70.6% were still cited. Blocking OpenAI's search index bot, 82.4% were still cited. Blocking the training crawler GPTBot, 88.2%. Blocking Google-Extended, 92.3%. Two named cases make it concrete: cnbc.com blocks three OpenAI agents at once and appeared 1,298 times in the dataset, and yahoo.com blocks Google-Extended and appeared roughly 30,000 times.

Google-Extended in particular is widely misunderstood. It is a robots.txt token, not a crawler with its own user agent, and it governs whether already-indexed pages may be used for Gemini training and grounding. It does not remove you from AI Overviews or AI Mode. There is no Gemini crawler at all, because Google grounds Gemini on content fetched by ordinary Googlebot.

The failure worth worrying about is the opposite one, and it is invisible. A firewall or content delivery network quietly returning a 403 to legitimate AI retrieval bots is far more common than a deliberate block, and you will never see it in robots.txt. Check your server logs, and make the blocking decision on the merits rather than as a reaction.

What to do this week

Write one page. Three headings: what I paid for, what I received, what I could not verify.

The third column is your specification for the next engagement. Every item in it becomes a required deliverable, a required access, or a required report line.

Then check your four logins. Registrar, profile, analytics, site admin. Fix whatever you cannot access, today, regardless of what you decide about hiring.

Then take the hour and write the six baseline numbers down, even if you have no intention of hiring anyone for six months. The value of a baseline is that it exists before you need it, and it is the one protection on this list that does not require anyone else to agree to anything.

When you are ready to look again, start narrow. Two or three candidates, references from businesses like yours, and the same five questions to each. The process for finding a good search company works better the second time, because you now know exactly which failure you are screening for.

Questions about what actually went wrong last time?

Email me at eric@seod.com with a short description of what you bought and what you got, and I will tell you which of the four failure modes it looks like and the two contract clauses that would have prevented it. That is the whole reply. I am not asking to be the next agency, and there is no message after it.

There is more on hiring and working with agencies when you are ready to start looking.