In Google’s AI Mode, 95 of every 100 queries end without a click. In ChatGPT, it’s somewhere between 78 and 99. The traffic isn’t dying — it’s relocating. It now lives inside the answer itself, in the paragraph the model writes about you (or doesn’t).
That’s the whole reason GEO exists. The goal stops being “rank my page” and starts being “show up in the answer, named correctly, with the framing I want.” Below is the playbook — the moves that actually shift that, the ones that look productive but don’t, and how to measure either honestly enough to argue with the CFO.
The two systems that decide whether you’re cited
Every AI surface is one of two things, and the move that works on one is wasted on the other.
- Foundation-model answers. The model leans on what it learned during training. If you’re not already a known entity in that snapshot, no amount of new content this quarter will surface you. ChatGPT without browsing, Claude without tools, the older Gemini fast paths — all foundation-mode. Influence here is long, slow, brand-presence work: be everywhere a crawler scraped the web, repeatedly, in stable language.
- Retrieval-augmented (RAG) answers. The model runs live searches before answering. Perplexity, ChatGPT browsing, Google AI Overviews, Gemini grounded mode. Influence here is fast and tactical: the corpus the model retrieves from is the same indexable web you’ve been working on for twenty years, just queried differently.
Most “GEO is a 90-day strategy” promises only make sense for RAG. The foundation-model side is a year-plus job. Sequence accordingly.
What to actually do
Four moves, ranked by how much they shift share-of-answer in our prompt panels. Everything else is rounding error.
1. Get cited on the sources the models already trust
Across our four-model panels, the same handful of domains keep appearing as citations: reddit.com, the obvious comparison sites (G2, Capterra, TrustRadius), the big editorial outlets in your category, and a long tail of category-specific roundups. Reddit alone often clears every other domain by 3×.
The move: list the top 20 cited domains for your buying-stage prompts, then go get on them. For Reddit, that’s genuine answers from a real account on threads where you have unique experience. For G2/Capterra, it’s a complete profile, recent reviews, and answered Q&A. For roundups, it’s outreach with a sharper pitch than the agency template everyone else is using.
This is the highest-leverage GEO move there is. Almost everything else is a multiplier on top of it.
2. Make your own pages legible to a model, not just a reader
Models lift sentences. If your homepage opens with “We’re transforming the future of work”, no model can paraphrase it into a useful answer about your product. They lift the sentence that is the answer.
What works:
- A one-line definition of what you do, in plain language, in the first 100 words. (“BrandAxis tracks how brands appear in AI-generated answers across ChatGPT, Claude, Gemini, Perplexity, and Grok.”)
- Comparison tables with named competitors. Models pull these directly.
- FAQ sections written as the question users actually type, not the question your SEO tool suggested.
- One H1, clean H2/H3 hierarchy, schema where it earns its keep (
Product,FAQPage,Article).
What’s overrated: heroic schema buildouts, every-page TL;DR boxes, “AI-friendly” formatting frameworks. They’re fine; they’re not the lever.
3. Fix the entity record before the model writes it for you
Every model carries an internal sense of what your company is. If that record is wrong — outdated description, wrong category, prior name, an acquisition the model missed — the answer about you will be wrong, and no on-page change fixes it.
Audit it once: ask each model “what is [your company]?” and “what does [your company] do?” with no other context. Note every error. Then chase down the upstream sources — Wikipedia, Crunchbase, your own About page, third-party profiles — that are likely feeding the model the bad version. Updating those is what eventually changes the answer. Expect a months-long lag on foundation-mode models; faster on RAG.
4. Defend recency
Models weight fresh sources heavily on RAG queries — Perplexity sometimes pulls >50% of citations from the current year. A two-year-old blog post can quietly become the answer about your company forever, even after you launch a better one, if no one is updating the corpus around you.
The fix is unglamorous: republish your foundational content with a current date and a real update note, refresh the comparison and pricing pages quarterly, and make sure new product launches generate at least one third-party article in the same week. Recency is not gaming — it’s the cost of staying the canonical answer.
What to skip
The GEO content economy is full of advice that sounds smart and changes nothing. The biggest offenders:
- Stuffing every page with FAQ schema. Helps your snippet eligibility on Google. Doesn’t move citation share on ChatGPT or Claude. Do it once, move on.
- Rewriting every heading as a question. Marginal. Mainly useful on long-form pieces where the question form genuinely matches a search query.
- Author-bio buildouts on every blog post. Models don’t yet weight author E-E-A-T the way SEO professionals hope they do. Worth doing for the 5–10 cornerstone pieces; not worth doing site-wide.
- “Track 25–100 prompts daily.” You don’t need that many to know what’s happening. 20 carefully chosen prompts that cover your buying journey, tracked weekly across 3+ models, gives you the same signal at 1/10 the noise.
- Generic “use lists, write clearly” advice. It’s just good writing. It’s not GEO.
What to measure
Three numbers, weekly, segmented by model:
- Share of answer. Of the prompts you track, what percentage name your brand at all? This is your headline metric. Most categories sit between 5% and 40% for the leader; everyone else is fighting for low single digits.
- Position in answer. Named first, named in the middle, or named in a “competitors include…” tail. First-mention is roughly 3× as valuable as a tail mention by every downstream metric we’ve measured (clicks, branded search lift, demo requests).
- Citation source mix. Of the URLs the model cited when it mentioned you, where did they come from? If 80% of your citations are your own domain, you’re in a fragile position — one model update can erase you. Healthy mixes lean heavily on third-party sources you don’t control.
Sentiment is a fourth metric some people track. Useful, hard to automate well; treat it as a quarterly read, not a weekly one.
What not to measure: raw “AI traffic” in GA4 alone. Most AI clicks come through with referrers stripped or aliased; you’ll under-count by 40–60%. Use it directionally, never as the primary number.
A 30-day plan that fits on a Post-it
- Week 1. Pick 20 prompts that match how your buyers actually ask. Run them across ChatGPT, Google AI Overviews, and Perplexity. Record share of answer, position, and the cited source list. This is your baseline.
- Week 2. Audit the entity record. Fix Wikipedia, Crunchbase, your About page, and any out-of-date third-party profile in the citation list. Rewrite the first 100 words of your homepage to be model-liftable.
- Week 3. Identify the top 5 third-party domains showing up in your category’s citations. Pick the two where you have the strongest case and go earn placement — review, op-ed, Reddit answer, comparison feature, whatever fits.
- Week 4. Re-run the same 20 prompts. Compare. The honest answer in month one is usually “RAG moved a little, foundation-mode hasn’t moved at all.” That’s the right outcome — keep going.
Where GEO doesn’t pay off (honest version)
A guide that won’t tell you this is selling you something:
- Sub-$500 ACV B2C with no brand search. AI answers route to brands the model already knows; if buyers don’t search you by name yet, GEO compounds slowly.
- Highly regulated categories (medical, legal, financial advice). Models hedge or refuse. Share of answer numbers will be misleadingly low across the board.
- Pure brand-defense plays where the model is already favorable. If you’re the category leader and every model already names you first, more GEO investment hits diminishing returns fast. Spend on the next thing.
- Zero-distinction products. If a model can’t articulate why someone would pick you over the named competitor in the same answer, no amount of citation-building will fix that. Fix the product story first.
GEO is a lever, not a religion. Apply it where it has leverage.
The takeaway
The brands winning AI answers right now aren’t the ones running the most elaborate GEO programs. They’re the ones who picked the four moves above, did them with discipline, and ignored the noise. Pick your 20 prompts this week. Run them. Decide one move to make. That’s the whole game.