GEO vs SEO: what changes when AI assistants answer the question

A customer types a whole situation into an assistant: "We run a five-person agency on QuickBooks and need simple project tracking for under $30 a user. What should we look at?" A few seconds later they have a paragraph, a shortlist and a handful of small numbered links. Your page might be one of those links. It might have been read and paraphrased without credit. It might never have been fetched at all.

So founders ask a fair question. Is generative engine optimization (GEO) a new discipline that needs its own budget, or is it SEO in a new coat?

Mostly the second. The machinery under every major assistant is still a search index, and the work that gets a page found there is the same work it always was. Some things do change. The unit that gets used is now a passage, where it used to be a page. The questions are longer and more specific. A visit, when one arrives, means something different. Those changes are worth adjusting for. They are not worth a separate department.

How AI assistants read the web

Strip away the interface and the major assistants follow the same routine: search, read, summarize, cite.

Google describes AI Overviews and AI Mode as grounded in its Search index. The technical name is retrieval-augmented generation, which means the model looks things up before it writes instead of answering from memory alone. Google also uses what it calls query fan-out: one complicated question is split into several smaller searches that run at once. The agency prompt above might become separate lookups for project tracking tools with a QuickBooks integration, per-user pricing for small teams and simple tools built for agencies. Each sub-search returns pages. The answer is assembled from pieces of them.

OpenAI says ChatGPT search rewrites the user's prompt into one or more targeted queries, sends them to its search providers, reads the pages that come back and attaches inline citations that link to the sources. Perplexity calls itself an answer engine, combining live web retrieval with a language model and numbering its citations in the text. It has also written about judging extraction at the level of spans, favoring precise, relevant passages over long generic blocks. Microsoft's documentation says Copilot generates queries to the Bing index to ground its answers in current web data, and in workplace versions it can blend those results with an organization's own data through Microsoft Graph and Copilot connectors.

Every one of these begins with a page that a search system can find, crawl and understand. If yours fails that test, nothing downstream rescues it.

What traditional SEO still handles

Google is unusually direct on this point. Its guidance for AI features says the foundations of SEO still apply: pages need to be indexed, offer a sound page experience and be genuinely helpful to the people reading them. Google also says there is nothing extra to buy, submit or install to appear in AI Overviews. No separate index, no secret protocol.

That makes the dull checklist the important one:

• Pages return a 200 status, load properly on mobile and are not accidentally blocked in robots.txt.
• Each important page is indexed. Search Console will tell you.
• Titles and headings describe what the page actually says.
• Internal links connect related articles, grouped into topic clusters, so a crawler can find the specific page as easily as the broad one.

One addition many small teams skip: Bing. Microsoft says Copilot grounds its answers in the Bing index, which makes verifying your site in Bing Webmaster Tools a small job with an obvious payoff. Online shops have a further input. Product data submitted through Google Merchant Center, including attributes such as product highlights and product details, is structured information Google can draw on in its AI-driven surfaces. Keeping that feed complete and accurate is unglamorous and useful.

A GEO plan that skips all of this is a roof with no walls under it.

What generative discovery changes

The differences show up when you compare the two across the same attributes.

Take the query first. Keyword tools measure short phrases, and short phrases are not how people ask assistants for help. A long, specific prompt has no search volume figure, yet it is exactly the kind of question your sales team hears every week. Keep using keyword data to choose topics. Then write each article around the real scenarios inside the topic: the team size, the existing software, the budget ceiling, the objection.

Next, the click. A page that explains what a chart of accounts is may be read, used in an answer and never visited, because the reader's curiosity was satisfied above the links. That is uncomfortable, and no amount of formatting reverses it. The click that does arrive from a cited source tends to come from someone who wanted to verify a detail before acting. Fewer visits, and plausibly better ones.

Then, originality. A model can summarize ten near-identical articles about choosing accounting software without citing any of them, because none of them says anything the others do not. It needs a source when the answer depends on a fact that lives in one place: your measured setup time, your actual return rate on a product line, the exact steps your support team uses to fix a sync error. The generic 800-word summary was always weak. Now it is also the easiest page for an assistant to skip.

How to structure articles for AI citation

Nobody outside these companies can promise that a page will be cited, and answers vary from one run to the next. What you can do is make your pages easy to extract from and worth extracting.

Write passages that stand on their own. Put the answer in the first sentence under each heading, then explain. If a heading asks how long a QuickBooks integration takes to set up, the next sentence should give the time, with conditions, before any background.

Define things plainly. "A seat minimum is the smallest number of paid users a plan allows" can be lifted into an answer intact. "When it comes to seats, there are many factors to consider" cannot be lifted into anything.

Name the subject in every paragraph. Pronouns that point three paragraphs back, such as "it" or "this solution", make sense to a reader scrolling in order. They make no sense to a system that has pulled out one paragraph alone.

Use a table when readers compare options on the same attributes. Three project tools set against price per user, QuickBooks sync and seat minimum give an assistant clean, parseable facts. A warm opening paragraph about today's fast-moving business landscape gives it nothing to quote, and it will not try.

Add what only you know, and only what is true. Your own pricing, your own process, your own figures with the date you measured them. Invented benchmarks are worse than none: they are easy to repeat and hard to retract.

Check every source by opening it. A citation to a page that has moved, or that does not say what you claim, undermines the trust you are trying to earn. Some tools build this in. Rune, an AI blog writer for brands, fetches every URL it cites before the writer uses it and drops any that fail to load. A person with a browser tab and some patience can do the same check.

On structured data, Google says its AI features need no special technical requirements beyond the basics. Standard markup that matches what is visible on the page, such as Product, Article or Organization, still helps search systems understand what the page is. Treat it as housekeeping, not strategy.

Evaluating performance in a zero-click world

If informational traffic falls while impressions hold steady, the content has not necessarily failed. It may be doing its job somewhere you cannot see.

Search Console's Performance report is still the first place to look, page by page. Watch the shape of each line instead of the site total. On a definitional article, impressions holding while clicks drift down is consistent with a question being answered before anyone reaches the links.

In your analytics, look for referrals from assistant domains such as chatgpt.com, perplexity.ai and copilot.microsoft.com. Volumes will usually be small. Look at what those visitors do: demo requests, contact forms, add-to-cart, time on a pricing page. That is where the evaluation-stage value should show up, if it shows up at all.

Then judge articles by their job. A definitional piece earns its keep by holding impressions and linking readers to the pages that convert. A comparison or buying guide should be judged on leads and sales. Once a month, ask the major assistants the five questions your customers ask most and note whether you are cited and who is. Treat it as a sample, since answers shift between runs. It still tells you which competitors supply the facts in your category, and which facts you are missing.

So spend the GEO budget on the SEO you have been putting off, then on articles only your business could write. Fix crawling and indexing, verify your site with Bing, keep product data clean and give every important page a passage that answers a real customer's question in its first sentence. None of this guarantees a citation, and nobody honest will say otherwise. It does make your pages the kind assistants can use. If drafting time is what stops you, Rune's free preview writes two illustrated articles from your site and a short brief in about 5 to 15 minutes, which is enough to judge whether its drafts meet your standard before you commit to anything.