Last week a friend running a cross-border SaaS product sent a question. He typed his category into ChatGPT, asked which options in this space were worth a look, and the answer named two competitors and not him. Nobody on the team was surprised; the product is newer. The surprise was the next message. Someone was already pitching him a GEO package with a three-month guarantee of placement in AI recommendations.
Conversations like this get more common every month. Generative search has moved part of the attention that used to sit on the results page into AI answers. For a brand, whether you show up in those answers has become a question people ask out loud. The services and claims around it, though, are far less mature than the SEO industry was at a comparable age, and even the search platforms publicly say that a guarantee of placement cannot be made.
The conclusions first. One, the major platforms have published part of how generative engines pick sources, and that part is enough to guide everyday work; the rest is a black box, and promises of insider metrics deserve no trust. Two, the foundation of this work is still crawlable, indexable, trustworthy, understandable; what changed is the unit of competition, from page ranking to paragraph-level citability. Three, citations can be observed, but no current method gives an exact answer, and the three common approaches each carry blind spots. The workable stance is to run observations with the limits made explicit.
First, the Published Part of How Answers Pick Sources
Start with the platform that reaches the most people. Google keeps a guide in its search developer documentation devoted to generative AI search, and it spells out the machinery more concretely than outsiders expected.
Its AI Overviews and AI Mode pull sources through two paths that run at once. One is retrieval-augmented generation, where the system uses core Search ranking systems to find relevant, recent pages from the index, then has the model compose an answer from what those pages actually say, with clickable source links attached. The other is query fan-out, where the model breaks a complex question into several related sub-queries and runs them in parallel, so the answer typically draws on a wider set of sources than a single classic search would. Who gets into the retrieved pool is decided by ranking and quality systems, the same machinery classic SEO works with.
The eligibility line is equally blunt in the documentation. To be shown as a supporting link in an AI answer, a page must be indexed and eligible to appear in Search with a snippet, and that is all. No additional technical requirements, no special files created for AI readers, no special structured data. The reverse holds too. Meeting every requirement does not guarantee a page is crawled, indexed, or served; the documentation puts it as not guaranteed. That sentence is worth rereading, because it explains why any service promising guaranteed citations can be dismissed.
Other platforms each run their own entry rules, and the publicly confirmable part sits at the crawler layer. OpenAI's documentation splits its visits into identities. GPTBot handles the crawling that points toward model training. OAI-SearchBot determines whether content can surface in ChatGPT's search features. ChatGPT-User represents a person fetching a page while asking a question. All three are controlled separately through robots.txt, and a change takes about a day to take effect. Perplexity works the same way, with PerplexityBot serving its search results and Perplexity-User handling visits a user triggers on the spot. These settings decide what enters the candidate pool. To check whether a platform has been dropping by, look for the matching user agent in your server logs.
The black box does not need our help either. Weights, the way an answer composer settles on sources, differences between platforms, none of it is published. Any tool or course claiming access to AI internal metrics deserves suspicion, and Google's guide says as much, noting that no third-party tool can see its internal ranking or AI systems. The workable attitude is this. Published mechanics guide the work; the rest gets approached through observation, not through faith in some secret recipe.
Generative Engine Optimization Rests on the Same Foundation
Treating it as a brand-new discipline leads people off the path. Retrieval, ranking, trust evaluation, the intermediate steps a generative answer depends on, are the same parts the search system always had. The difference lands in how the answer gets assembled. Classic search hands you a line of results and the user digs through the page. A generative engine extracts specific information from the page and folds it into the reply, and the unit it takes is the paragraph, not the page.
One inference follows directly. Optimization used to target where this page ranks for this term. Now there is another layer on top, whether this page contains a stretch of text that can be lifted as an answer on its own. The two do not conflict. Ranking decides whether you enter the candidate pool; once inside, whether your sentences get taken decides your odds of appearing in the answer.
This is also the most interesting piece of the official stance. Google's guide answers the trendy terms AEO and GEO head-on, and the gist is that from Search's point of view, optimizing for generative AI search is optimizing for the search experience, which remains SEO. In other words, this stays part of the SEO job, and there is no need to treat it as a new discipline.
The Actions With Evidence, From Foundation to Citability
The priority actions in the official guide overlap heavily with classic SEO. Content that can be crawled, pages discoverable through internal links, a good page experience, important information presented as text, structured data that matches what people see, images and video handled under existing guidance. The guide also singles out a kind of content it values. Pages carrying first-hand experience and a distinctive point of view fall into what it calls non-commodity content, and they stand a better chance of being drawn into generative experiences than listicles restating common knowledge. The don't-bother list is just as clear. No llms.txt or similar special files built for AI, no chopping content into tiny pieces, no rewriting the voice to please AI systems, no buying artificial mentions across the web. Mass-produced low-value content walks straight into the platform's anti-abuse policies.
Academic work offers one more thread worth following. The GEO benchmark study, first posted in late 2023 and revised in 2024, compared a range of rewriting tactics in a simulated environment. Adding source citations, quotations, and statistics lifted content visibility in generative replies by as much as about forty percent, with results varying by domain, while keyword stuffing produced no gains. Label this clearly. Those results come from a simulated benchmark, not any platform's real mechanics, but they line up with the official preference for non-commodity, traceable content.
For a weekly checklist, this is workable. Under every core topic, keep one definition or conclusion sentence that stands on its own outside the page context. Attach a source and a year to every key figure, and leave out numbers with muddy definitions. Put author, organization, and update date where people can see them, not only in metadata. Reserve question-and-answer structure for real questions instead of stenciling it across the site. Consolidate duplicate content so the same answer does not appear in several diluted versions.
These actions share a trait. None of them were invented for AI, and all of them move content closer to a state where it can be lifted and checked. Doing them guarantees no citation either, only better odds.
How to Know Whether You Have Been Cited
Start from a premise. No tool today produces a complete list of citations. Three approaches are common, and each has blind spots.
Platform panels. Search Console already reports performance data tied to generative AI, and both its definitions and coverage keep evolving; the details and uncertainty were tracked in our September piece. Its limits are seeing one platform only, and not separating citation from click.
Referral traffic and logs. Server logs showing visits from OAI-SearchBot, ChatGPT-User, or PerplexityBot answer whether the platform has been around. Analytics sessions arriving from chatgpt.com, perplexity.ai, and similar sources answer how many visitors came with a link in hand. Both numbers run far lower than reality, because most citations never produce a click. Pew Research Center published a browsing study in July 2025. On search pages showing an AI summary, users clicked a traditional result link in about eight percent of visits, against about fifteen percent without one, and clicked a link inside the summary in just one percent. Clicks shrink, which is exactly why citations deserve their own observation.
Spot checks. Pick ten to twenty questions tied directly to your business, ask them on the major AI platforms once a month in fixed wording, and record whether the brand appears and whether a link is attached. The value is in comparing trends. The limits are just as clear. The sample is not random, models change, and results do not reproduce, so it suits internal reference rather than settlement.
Stitched together, the three give a sense of direction, not a precise account. Citations and rankings are two different ledgers, and we have already written up the measured data on this, so this piece will not repeat it. One line remains. A finer measurement design belongs in its own follow-up.
The Order of Operations
If work starts this week, the sequence that makes sense runs like this. Check the foundation first, walking through crawling, indexing, speed, and structured data; there is a complete engineering checklist on the site. Then rework content, beginning with the dozen or so pages closest to revenue, filling in definition sentences, data sources, and author information. Most of this is rearrangement and supplementation, not rewriting. Next, confirm the access layer, checking that robots.txt has not fenced off AI user agents and that requests show in the logs; if they are entirely absent, solve that first. Finally, run the three observations, reading trends quarterly rather than weekly. Long-term content stewardship lives in the content strategy piece, and multi-market scenarios sit in the international guide; neither belongs in this article.
Common Misconceptions
Treating GEO as mysticism and buying the recipe. llms.txt, content chunking, a special writing voice, the official guide names each as unnecessary; tools claiming insider metrics cannot be trusted. That budget does more for you in the foundation and the content.
Using citations as a settlement metric. Every one of the three methods has blind spots, and single spot checks swing widely. As a settlement figure it distorts; as a direction check it earns its place.
Mass-producing content to fill slots. The platform's anti-abuse policies exist for exactly that, and undifferentiated content does not solve the credibility question while adding risk.
Dropping classic SEO to chase the increment. The foundation comes first. A page that cannot be indexed has no path to being cited.