How Do I Do Generative Engine Optimization (GEO) for My Website?

Generative engine optimization is three jobs: measure whether models mention you, fix the pages they should be able to lift, and earn mentions on the third-party sources they already trust. You can run a real program in 30 days without a new "GEO department." If you want the definition first, read what GEO is. This post is the operating system.

SEO still matters. GEO does not replace rankings. It adds a second scoreboard: citations in generated answers. Teams that only chase blue links will watch Overviews shift toward product-page citations and wonder why another listicle did not help.

The three jobs, without the manifesto

1. Measure. Track a frozen prompt list across ChatGPT, Perplexity, Gemini, Claude, Copilot, and Google AI Overviews. Use first-party reports where they exist. Google's generative AI performance reports cover Overviews, AI Mode, and Discover impressions. Bing's AI Performance covers Copilot citations. Everything else is a tracker.

2. Fix the source material. Models cite pages they can parse. Answer block under the H1, consistent entities, tables, dates, schema that matches visible text. Google's AI features docs: no extra requirements beyond indexed and snippet-eligible. The AI optimization guide is the same bar in more words.

3. Earn presence where models look. Reviews, comparison sites, docs, forums, and publications already in the citation set. Your blog is not the only corpus.

Skip any one job and the other two stall. Measurement without edits is a dashboard subscription. Edits without measurement are SEO cosplay. Third-party work without a prompt list is PR without a target.

Days 1–7: baseline, do not "optimize"

Write 25–40 prompts from GSC commercial queries, sales language, and "vs / alternatives" questions. Freeze the wording.

Run them once across the six engines. Log mention, citation, URL, competing domains, and a screenshot. Turn on GSC generative reports and Bing AI Performance the same week.

Allow OAI-SearchBot if you want ChatGPT Search eligibility. Confirm it in logs. Blocking GPTBot is a training decision, not a Search decision.

Do not rewrite the site in week one. You have no before. We said this in what GEO is and it is still the most ignored step.

Days 8–21: fix pages that already should have won

Sort losses by revenue proxy, not by volume. A mid-volume comparison you lose everywhere beats a high-volume definition you will never take from Wikipedia.

For each loss, classify:

Unextractable — you have the URL; the claim is buried. Edit.
Missing — no page holds the claim. Write one sharp page, not a pillar.
Off-site — the model prefers a G2, docs, or affiliate URL. Earn that mention.
Noise — one engine, one day. Wait for a second observation.

Edits that actually move citations:

Lead with the answer the model already tried to give.
Put numbers and compatibility in a table.
Kill contradictory claims across pricing, docs, and blog.
Date anything that goes stale.
Match structured data to the text on the page.

Do not invent GPT-5 ranking factors. OpenAI has not published them. Do not ship 20 AI-written posts because a tool called them "GEO content." Our ranking notes treat actionability as a specific fix, not a writer.

Days 22–30: re-measure and pick the off-site fights

Re-run the same prompts. Keep a before/after screenshot on anything you will claim in a meeting.

If the on-site edit did not move a three-engine loss, stop writing variants. Look at the cited domains. That list is the real corpus. Pitch, update, or advertise there. GEO is often closer to digital PR than to another cluster.

Then lock the operating rhythm:

Daily alerts on money prompts.
Weekly review of multi-engine losses.
Monthly prompt-set hygiene (add sales questions, kill vanity).
Quarterly competitor-set review.

After GSC and Bing are in the weekly ritual, the re-measure loop still needs a daily prompt portfolio for ChatGPT and Perplexity. Profound is the heavy version of that loop. Promptwatch is the one we put next to it in the tools list for teams that want six engines, screenshots, and alerts without an enterprise onboarding — measure, fix, re-measure, no writer bundled in.

What "good" looks like in 30 days

You will not "win AI search." You will have:

A prompt list leadership has seen and agreed to.
A baseline and a second measurement.
Five URLs that are actually extractable.
A short list of third-party sources the models already use.
A decision on which tracker owns the scoreboard.

That is a program. A sitewide visibility percentage is not.

How this sits next to SEO

Keep doing the SEO work that makes pages eligible: indexation, internal links, people-first content. Google has been consistent — AI features do not add a second technical bar. GEO adds monitoring and extractability on top.

When classic impressions hold and clicks fall on informational queries, you are often looking at the Promptwatch Overview pattern: listicles down from about 26% of citations in Q1 to about 18% in July, product pages leading daily from July 28. The fix is citation, not a title-tag pass. When you are uncited on a comparison prompt across chat engines, the fix is the claim and the source set, not another 2,000 words.

Use Ahrefs and Semrush for demand. Use GSC and Bing for first-party generative presence. Use a tracker for the prompt log. Use editors (or Goodie / Writesonic, judged as writers) for the fix. Do not buy one SKU that claims to be all four unless you have verified all four.

FAQ

Is GEO different from AEO / LLMO / GAIO?

Same job, different acronyms. We use GEO because that is what this site ranks. The work is citations in generated answers. See what GEO is.

Do I need a new CMS or schema product to "do GEO"?

No. You need crawlable pages, a prompt list, and a way to re-check answers. Schema that matches visible text is enough. A new CMS is a stall.

How is this different from content marketing?

Content marketing ships inventory. GEO decides which claims must be true, visible, and preferred versus a cited competitor. Most weeks you edit, not publish.

Can we skip third-party mentions and just improve our site?

Sometimes. When the model already likes your domain, extractability is enough. When it consistently cites G2 or a docs subdomain you do not own, on-site work will not steal the slot.

Should we opt out of AI Overviews?

Only for a legal or brand reason. Opting out removes generative impressions and does not restore old CTR. Most commercial sites want in, then want the citation.

What to do this week

Read what GEO is if anyone on the team still thinks this is "SEO for chatbots" as a slogan instead of a scoreboard.
Freeze 25 prompts. Run them once. Turn on GSC generative reports and Bing AI Performance.
Allow and verify OAI-SearchBot.
Pick the five losses that map to revenue and have an existing URL. Fix extractability only.
Put the 25 prompts on a daily tracker from the tools list. Re-run on day 21. Do not expand the content calendar until that loop exists.