articles · digital growth
How to choose a digital agency, and the flags that say walk (AI included).
Full disclosure before a word of advice: we're a vendor writing about vetting vendors. Read accordingly, and note that every question below works on us too, which is rather the point. Bad agencies survive because owners don't know what to ask until the retainer has already taught them, and AI has made it cheaper than ever to look busy. Here are the questions that teach it beforehand, including the ones worth asking now that more of the work is done by software.
skeelx · updated oct 2026 · 5 min read
First check whether you should be hiring at all: the DIY-or-agency split is its own decision, and "delegated" is only one of its three honest outcomes. If you land there, what follows is the interview.
The guarantee flag
Anyone who guarantees rankings is lying about how the machine works. Google's results are an auction of relevance run by someone else. No vendor controls them, including vendors with "partner" badges, including us. The honest version of ambition sounds different: here's the work, here's the mechanism it feeds, here's how we'll measure whether it's paying (the crossover arithmetic is what that measurement looks like). "Page one in ninety days or your money back" isn't confidence. It's a funnel for people who want certainty more than results, and refund terms designed never to trigger. Its newest form is a guaranteed place in AI answers; those engines belong to someone else too, so the same reasoning applies. Guarantee talk on the first call saves you the second one.
The hostage test
The single most expensive discovery owners make about a bad agency arrives at the end: the ad account was "theirs", the website lives on "their platform", the analytics belong to "their agency account", and leaving means starting from zero. Years of campaign learning, conversion history and content, held as leverage against your departure. So ask it straight, before anything is signed: who owns every account, and what exactly do I keep if we part on bad terms? The only acceptable answer is that everything (ads, analytics, the website, the domain, the business profile, the CRM) sits in accounts your business owns, with the agency working as an invited user. This one question outranks every other on the list. If the answer is wrong, nothing else on the card matters.
"What did you ship last month?"
Retainers rot when activity replaces output. The question that exposes it is concrete: for a current client of similar size, what actually shipped last month: which pages went live, which tests ran, which campaign changes were made and why? A working agency answers instantly, because the list exists; it's how they run the work. A rotting one answers with process ("we run continuous optimisation") or offers to show you a dashboard. AI makes the old trick easier: a month of generated posts and pages can pass for output, so ask what each item was for and what it changed. Reports describe weather. Shipping is what you're paying for, so buy from people who can name what they shipped.
Reporting in money, not rankings
Ask to see a real (anonymised) monthly report before signing, and read it for one thing: what's the headline unit? Impressions, sessions and "visibility scores" are activity dressed as achievement. The report you want leads with enquiries, the customers they became and their cost (which channels produced them, what each cost, what changed since last month and what will change next) in language a business owner can act on. An agency that reports in rankings is negotiating with your attention; an agency that reports in money is accountable to it. The difference tells you what they'll optimise for all year.
The agent-native lens: your customers' assistants, the agency's agents
When your customers send assistants. Some buyers now hand the early research to an assistant. Ask every shortlisted agency how they would make your business easy for that software to read and check, and listen for specifics rather than a new acronym: facts that agree across your site, profile and listings; schema.org JSON-LD that matches what the page actually says; pages that answer the questions buyers ask; and, where it suits the business, a way for software to check availability or book without a phone call. Then ask how they'd know whether it was working. A good answer admits the measurement is still young and proposes something concrete to watch.
When the agency runs on agents. Plenty of agencies now use AI to draft, monitor and report, and that can be good value; it just adds a layer to the interview. Which work is done by agents, and which person reviews it before it goes live or spends your money? What access do those agents have to your accounts, and is it the narrowest that works? Are their actions logged under an identifiable user you can see? And the hostage test, extended: if they build automations, prompts or agent workflows on your accounts, are those documented and yours when you leave? An agency that answers plainly has thought it through. One that can't say what its software did last month has the old activity-without-output problem, running faster.
Contracts, and the reference call
Contract terms tell the truth about confidence: month-to-month or short notice periods mean the work is expected to justify itself; long lock-ins mean the paperwork is doing that job instead. Read the exit clause first: ownership on departure, handover obligations, and any fee for leaving. Then make one reference call, to a client who's been aboard long enough for the honeymoon to be over, and ask two questions: what do they actually do each month, and what happened when something went wrong? The second answer is the audit. Every agency performs well in the sales cycle; references describe the marriage. The tools agencies use are likely to keep changing faster than the questions worth asking them: who owns it, what shipped, what it earned, how we leave.