Skip to content

A practical AI workflow for local SEO research and client reporting

AI tends to be most valuable at the outset of a piece of work rather than at its close.

This guide sets out the working approach of Dana DiTomaso, founder of the agency Kick Point and creator of the Analytics Playbook training platform, in a form others can follow: lean on AI to overcome the blank page, keep the judgement with a person, and produce reports that clients will genuinely open. The material is drawn from a conversation she took part in that has already been published; it is not a fresh interview, and it does not imply any endorsement.

Start with the measurement problem

For years, traffic was the shorthand for success. AI answers and LLM summaries have weakened that link, and no single replacement has settled yet. Two candidates are worth watching: citations, which can mean either a brand mention inside an AI answer or an AI crawler visiting your site, and branded search trends in Google Search Console. Treat branded search as a directional signal rather than a precise score; read the trend, not the number.

Underneath both sits a content problem. Years of investment in top-of-funnel pages left middle-funnel content — comparison and consideration material — thin. That gap is where an AI answer can step in with an answer your own site never gave. Local AI visibility trackers can show how you are mentioned when Search Console or Analytics fall short.

Find the content you are missing

The approach is intentionally straightforward and can be run again and again:

  1. Export roughly 90 days of Google Search Console query data as a CSV.
  2. Load that export into Claude together with your keyword research.
  3. Attach an SEO tool that supports MCP so ranking data sits in the same conversation. SE Ranking and the MCP server offered by the platform publisher are two examples; DataForSEO and SEMrush were named as further options.
  4. Have the model group queries by funnel stage and mark the areas that are underserved.
  5. Have it set priorities against the time you genuinely have, rather than an idealised team. If you work alone and can write for two hours on Tuesday mornings, state that and ask where to begin; the result becomes a plan you can carry out instead of a list that overwhelms you.

When the gap is identified, write for the way people really put questions to an AI. Nobody enters "budgeting software" into an LLM; they describe their circumstances — revenue, location, whether an accountant handles the books — and expect an answer shaped to them.

Build a reporting assistant that knows your standards

A reusable Claude project can act as a reporting brief that never forgets your house style. The setup described here began with a large body of accumulated guidance on Looker Studio report design: the formulas used, the layout conventions preferred, and the elements that should never appear in a client dashboard. Once that material is in place, supplying a client brief or an unfinished article produces a structured markdown outline. That outline covers the questions the intended audience will bring, a suggested arrangement of the dashboard, short explanatory text for individual charts, and the introductory page that opens every report. A worked example came from an Analytics Playbook article on whether Google Search Console impressions can be trusted and how to detect it when they cannot. The draft went in, and the response was a complete dashboard plan organised around that question, with a bar chart, a heat map and a calculated field among the proposals. The calculated field was dropped on the grounds that the same figures could be obtained from GA4 without it. That decision illustrates the value of the arrangement: the output is a starting proposal to be interrogated, not a verdict to be accepted, and pruning it prevents unnecessary work from accumulating. Building a dashboard from scratch is estimated to take a few hours less as a result. The workflow stops well short of automation, however. Claude and Looker Studio are not directly connected, so every chart still has to be assembled by hand; what has changed is how much thinking is needed before construction begins. The quality of the input material is as important as the tool. Business context is not optional, because without it an AI assistant will misread ordinary patterns. In one instance a question about whether sessions had risen or fallen was put to Claude with no seasonal background attached, and the omission was spotted mid-question: for a B2B business, Christmas and summer are quiet stretches, so a dip at those times means little. Context worth supplying includes transcripts of client calls, Notion pages recording significant decisions, and a brief onboarding questionnaire asking how the client's performance is judged and which chart most recently helped them reach a decision. Strip that away and the recommendations revert to generic advice.

Design reports around decisions, not metrics

The sprawling multi-screen dashboard, which Dana DiTomaso describes as something out of Star Trek, is not where she expects reporting to go. A tighter format is preferable. It would present a small set of measures — somewhere between three and six — that receive regular attention, alongside a self-service option through which clients can pose their own questions and get answers shaped by the agency's professional judgement. That judgement is central to the design. Take engagement rate as an example: if a client requested it, the system should decline and give a reason, because the metric does not carry meaningful information. The concept remains exploratory. Dana is investigating, together with Britney Muller, whether an MCP could put this capability directly in clients' hands, and for the time being she deliberately keeps Claude projects from being shared between separate organisations. The underlying concern is simple: a dashboard that nobody opens has failed, whatever it contains.

Use AI to get started, not to finish

Dana DiTomaso notes in passing that she has ADHD, and for her the most valuable thing AI offers is support with executive function — a nudge towards the first step when a task feels hard to begin. That thread runs through the whole discussion. The aim is not to have AI complete the work on your behalf. It is to carry you across the starting threshold, past the empty page, so that your own expertise can be spent on the decisions that genuinely require it.

Frequently asked questions

What should I feed an AI reporting assistant?

Client call transcripts, records of major decisions, and a short onboarding survey covering how the client is evaluated and which chart last helped them make a decision. Without that context the output is generic.

Does the AI build the Looker Studio dashboard for me?

No. It returns a markdown spec with the audience's questions, a proposed structure, chart microcopy and an about page. There is no direct connection between Claude and Looker Studio, so the charts are built manually.

How many metrics should a client dashboard show?

Aim for 3 to 6 metrics people check regularly, plus a self-serve layer where clients can ask their own questions with your judgement built in.

Is branded search in Search Console an accurate success metric?

Treat it as directional. It may not be highly accurate for a given business, but the trend is a useful signal — read the trend, not the number.