Skip to content

Using AI to Run Original Local SEO Research: A Practical Workflow

Ben Fisher, founder of Steady Demand, describes a way of working in which AI takes on the tedious parts of local SEO research while he keeps hold of the question, the method and the final wording.

This guide sets out that approach as a repeatable workflow: start from a real problem, measure the friction before automating it, let a model handle collection and arithmetic, and publish enough detail that other practitioners can re-run the test. The claims and examples below are drawn from Fisher's account of his own practice, not from an independent study. Fisher has spent close to three decades in search and was among the first three Google Diamond Product Experts in the US. He discussed his approach with Myles Anderson, co-founder and CEO of the publication that hosted the conversation. The related series also covers Noah Learner at Sterling Sky, Dana DiTomaso at Kick Point, and Mike Forgie, a local search and AI optimisation expert; Learner and DiTomaso focused on reporting, while Forgie described streamlining agency workflows with Claude Code.

Start with a problem, not a tool

Fisher's rule of thumb is that the tool choice comes last. A project should begin with a question that is currently unanswered or a task that is currently slow. Only then does it make sense to ask whether a model, a script or a small custom app is the right way to close the gap.

That framing suits local SEO, where plenty of confident claims have never been tested at scale. A workable research question is narrow enough to answer with data you can actually collect, and specific enough that the answer would change what you or your clients do.

Worked example: the uncertain location count

Fisher describes a large enterprise client that believed it had around 150 locations but could not confirm the figure. The task was not a research paper; it was a bounded data problem with a clear output.

His approach was to build a scraper that matched a list of business names, consolidate the results, and present them in a simple tracking app the client could follow. The tool was deliberately disposable: it existed to answer one question and would be retired once the answer was settled.

The lesson is about the economics of small custom tools. A temporary need that once required a developer for weeks can now be met in a single working session, which changes the threshold at which building something bespoke is worth it. The output is a confirmed list and a progress view, not a permanent product.

Measure the friction before you automate it

A second pattern is to quantify a slow routine before changing it. Fisher tracked the time he lost moving between his inbox and internal client dashboards. The measurement showed more than two hours a day were going into the back-and-forth. Bringing the dashboard data into the email environment recovered roughly an hour and a half per day.

The order matters: measure, then automate. Without a baseline you cannot tell whether a new tool saved time or simply moved the effort somewhere less visible. A short tracking period is usually enough to decide whether a fix is worth building.

For teams building their own reporting views, the platform's MCP server is one route to pulling dashboard data into the tools people already use.

Keep a home for the tools you build

Once several small tools exist, the practical problem shifts from building to remembering. Fisher's answer is a personal central hub: one place that holds the to-do list, useful links, the day's schedule and every tool that has been built. It keeps the collection usable as needs change and prevents the same utility being rebuilt because nobody remembered it existed.

Keep control of the process, let AI do the heavy lifting

The most important control point is the research framework. Fisher sets the direction and outlines the method first. AI then assists with the mechanical layer: connecting APIs, gathering data and working through the numbers. The final step is human: refining the findings into clear, natural language that readers can follow.

A simple way to apply this is to decide, before any prompt is written, which decisions must stay with a person and which tasks can be delegated. The checklist below is a starting point rather than a rule.

Human-versus-AI control checklist

Use this to scope a research project before building anything.

  • Keep with a person: the research question, the definition of the population being studied, the sampling logic, the interpretation of results and the final wording.
  • Delegate to AI: API connections, bulk data collection, repetitive matching and consolidation, and arithmetic over large tables.
  • Review together: any statistic that will be published, especially where a small change in method would change the headline.
  • Record for publication: the exact steps taken, the data sources and the raw data, so a reader can repeat the test.
  • Retire deliberately: note which tools were one-off and which are worth keeping in the hub.

Publish the methodology and the data

Fisher argues that being as open as possible about method and data is what builds trust, and he notes that few people in the industry do it. Publishing the steps and the underlying data lets other practitioners repeat the tests and check the conclusions for themselves.

There is a practical reason too. AI answers and the citations behind them are not fixed, so a study that others can re-run later holds more value than a single frozen snapshot. The benefits run in two directions: internally, a team can make decisions backed by evidence it can defend; externally, smaller agencies can use detailed, repeatable research to build authority and to guide clients with something more solid than opinion.

Getting started without taking unnecessary risk

Fisher admits he was hesitant at first. Hands-on practice is the fastest way to build confidence, but it is reasonable to reduce exposure first. A separate laptop for experimentation, and a VPN in front of that machine where the concern is serious, are simple precautions. Team capability matters too: prompting is becoming a baseline skill rather than a specialist one.

From there, the advice is to start small and expect breakage. The first project does not need to be polished; what matters is getting comfortable with the loop of idea, attempt, troubleshoot and improve. That loop is what makes the later research work possible. Fisher's own framing is blunt: anyone willing to try, and able to do a little troubleshooting, can get a long way.

Frequently asked questions

What counts as original research in local SEO?

A study that produces findings from data you collected or assembled yourself, with a method someone else could follow. The value comes from the question being specific and the method being documented, not from the size of the dataset.

When is it worth building a custom tool with AI?

When a task is repetitive, bounded and would otherwise take a person hours of manual work. A disposable tool that answers one question and is then retired can be entirely reasonable, especially for a single client problem.

What should stay under human control in an AI-assisted study?

The research question, the method, the interpretation of the numbers and the final published wording. AI is best used for the mechanical layer of collection and calculation.

Why publish the methodology and raw data?

So others can repeat the test and check the findings. That is especially useful in AI-influenced local search, where results and citations change and a study may need to be re-run later.