The 4-step roadmap to AI agents for Google Ads
robert-simpkins in search-engine-land, 2026-08-14, 5-minute read. A contributor piece arguing that most Google Ads teams are not ready for an agent and should not build one yet, written by someone who spent the past year building them for clients. It is the counterweight to groas in this spoke's paid corner: same automation axis, opposite recommendation.
The argument
The framing claim is that the readiness question comes before the capability question. "One of the biggest misconceptions surrounding AI is that it compensates for poor processes. In reality, it simply automates those processes faster." Everything else follows from that: the four steps are a prerequisite chain, not a menu, and they are held on agentic-adoption-ladder.
1. Foundation first — knowledge base and data. Document products, services, business rules, tone of voice, campaign structure and internal processes "in a format AI can understand," and connect the marketing data so the model can see the whole account. Named as the step teams skip because it is the least interesting. On the warehouse question the piece is deliberately indifferent: "Whether you use BigQuery or another centralized warehouse matters less than eliminating the silos."
2. Exhaust off-the-shelf before building. Export campaign data into ChatGPT or Claude and ask it to audit account structure, find wasted spend, surface search-term opportunities or review the shopping feed. Then wire the same tools to Google Ads, Google Analytics and Merchant Center through pre-built MCP connectors, which turns a weekly spreadsheet export into a live query against the account with business context held in a project or custom GPT. "For many organizations, this combination will deliver the majority of the value they'll ever need."
3. Build custom only when the requirements get specific. The examples are joins the ad platform cannot do — advertising performance against stock availability, pricing, margin and CRM data — plus continuous monitoring instead of prompt-and-wait, and approval workflows with controls. What developers add is named precisely: custom MCPs, guardrails, orchestration, scheduling and cost optimization, which is what separates "an interesting demo" from something dependable enough for daily use.
4. The constraint is people, not technology. Find the enthusiastic early adopters, give them room to experiment, then spread what works — rather than expecting the whole team to become expert at once.
What it refuses to claim
The closing section is the part worth keeping. "The biggest mistake I see is businesses trying to automate every aspect of Google Ads. That isn't where the value lies." Agentic AI is placed as good at repetitive, data-heavy work — auditing, monitoring, trend analysis, surfacing opportunities — and the marketer keeps strategy and judgment. The heading says it plainly: the goal isn't autonomous marketing.
That sits directly against groas, which sells a Google Ads agent as a replacement for the PPC manager. Both are practitioners with commercial skin in the game and they disagree about the ceiling, which is recorded as a contradiction in synthesis rather than resolved here.
The historical claim, and why it matters more than the roadmap
"We've already spent the last decade handing more execution to algorithms via Smart Bidding, broad match, and Performance Max. Agentic AI is simply the next stage of that evolution."
This is the sentence that earns the ingest. It says agentic PPC is not a break but the latest step in a delegation that advertisers have been consenting to for ten years, and it names the three platform-side steps that came before. paid-search-automation holds that arc. It also reframes the build-versus-buy question the roadmap answers: the agent an advertiser builds sits on top of an account whose bidding, matching and placement decisions were already delegated to Google, so what is being automated at step 3 is the supervision of an automated system, not the campaign.
Tier and standing caveats
T3. A contributor article with no data — no client count, no before-and-after, no measured outcome. Its authority is one practitioner's year of building these systems, and that practitioner co-owns an agency (Propel) that sells exactly the step-3 custom development the piece says most companies do not need. The self-interest cuts both ways here, which is unusual and worth noting: the advice most likely to cost him work is the advice to stop at step 2.
Two further things about the container. search-engine-land is owned by Semrush, disclosed at the foot of the article, and two Semrush promotions are embedded mid-text. And the piece is editorially processed — written by Simpkins, edited by Anna Sonnenberg, reviewed by Danny Goodwin — so it is a reviewed contributor column rather than an unfiltered guest post.
Related
paid-search-automation · agentic-adoption-ladder · robert-simpkins · search-engine-land · groas · ai-content-gap-analysis-workflow · seo-operating-model-shift · b2b-ppc-metrics · synthesis