How do I do generative engine optimization (GEO) for my website? Our client workflow
The GEO workflow we run for client websites: crawl access, indexing, prompt research, answer-first pages, off-site mentions and measurement, mapped to Promptwatch.
When a client asks how to do generative engine optimization (GEO) for their website, we start by lowering the temperature. GEO is not a separate discipline with secret levers. Google's guidance on AI features in Search points to the same fundamentals as ever: pages that can be crawled and indexed, with content that helps people. What changes is the measuring, and the set of places you need to show up.
This is the order we work in. We run it on Promptwatch, so the feature names below are from that platform.
1. Check access before anything else
We have seen programs spend a month on content while a CDN rule quietly blocked an AI crawler. So the first job is a bot access review. We read robots.txt and edge rules against the real list: GPTBot, OAI-SearchBot and ChatGPT-User, ClaudeBot, Claude-User and Claude-SearchBot, PerplexityBot and Perplexity-User, Google's agents, and the Meta and Mistral fetchers. OpenAI documents OAI-SearchBot as the crawler for search, so we treat that one as the one to allow if the client wants to appear in ChatGPT Search. Training bots are a separate policy decision, and the client makes it.
Then we confirm with logs rather than assumption. Promptwatch's Agent Analytics gives IP-verified crawler logs with spoofed user agents filtered out, error tracking with explanations, and a citation rate per page. Crawler logs start on the Professional plan.
2. Fix indexing and snippet eligibility
Pages need to return a 200, avoid stray noindex tags, put real text in the HTML and have sensible titles and headings. We check Google Search Console and Bing Webmaster Tools first. Promptwatch's Search Insights overlays Search Console and Bing data on AI crawls and citations, so we see where a page is indexed but never fetched by AI bots, or the reverse.
3. Build the prompt set
Keyword lists are a starting point, not the answer. We interview the client's sales and support teams, then write prompts the way customers talk: "best X for Y", "X versus Z", "how much does X cost", "is X worth it for a small team". Engines split these into sub-queries, so we also look at query fan-outs per model. In Promptwatch we use Prompt Explorer to pull suggestions from Search Console, ranking keywords and People Also Ask, and tag prompts by topic, type and intent. Personas matter too, since a founder and a procurement lead ask differently.
4. Write and refresh the pages
Our rule is to answer first, then support. State the point in the opening lines, then give evidence, specifics and limits. One topic per page. Real names, dates and numbers the client can stand behind. Structured data only where it honestly describes the page.
To pick which pages, we use content gap analysis, which scores coverage against the tracked prompts. For existing pages, Content Refresh in Promptwatch suggests diff-style edits. For net-new pages, the Content Agent (Business plan) can plan and write against gaps, with a review inbox or a hands-off mode with a quality check. We keep clients on the review inbox. Someone who knows the business should read every page before it goes live.
5. Earn presence off-site
Engines often cite sources other than the client's own domain. We open the citation reports and list which domains win the answers we care about: Reddit, YouTube, LinkedIn, comparison sites, trade press. Then we ask what the client can honestly contribute there. Promptwatch's offsite mentions view helps find pages that name the brand without linking. We do not manufacture reviews or astroturf threads. That tends to backfire and it is not advice we give.
6. Measure, and tie it to business
We report visibility score, share of voice, sentiment and position in answer, and compare periods before and after changes. More useful for the client's leadership is traffic and conversion: Promptwatch's visitor analytics counts AI-referred sessions and conversions via a lightweight script, a GTM template or the WordPress plugin. The Action Board then gives the team a prioritized list, and Agent Chat can answer questions about the data without exporting a spreadsheet.
What to do in week one
Check bot access, write 30 prompts, take a baseline, and pick the five pages that matter most to revenue. If you want a free first look, Promptwatch has an Explore tier (10 prompts, ChatGPT only) and free tools including an AI Visibility Report and an Agentic Readiness Check that need no login.
What we say about expectations
GEO does not guarantee a mention. Answers vary between runs and engines, and a small site in a crowded topic can take months to show movement, if it does at all. What you can do is remove technical blockers, publish better answers and measure honestly. That is the whole job, and it is enough to separate a client from competitors who are still guessing.