# Generative engine optimization for a personal brand Generative engine optimization for a personal brand, without the hype: what the research shows, what to publish, and which tactics put your name at risk. --- Generative engine optimization for a personal brand means making the true facts about you easy for an AI assistant to find, quote and check. In practice that is a short list of accurate pages on domains you control, written in sentences that stand alone, backed by independent sources that confirm them, and left open to the crawlers that feed AI search. It is not a way to make an assistant say something that is not true. The term comes from a 2023 research paper, and it has since become a label for a crowded market of advice. Most of that advice is written for products and companies. A person has a different problem: a name that may be shared, a career that changed, a few pages that matter and a reader who asks an assistant "who is this?" before a meeting. The work is smaller than the marketing suggests, and the risks are higher, because a tactic that looks clever on a brand site can read as manipulation when it is attached to a human being. This guide covers the optimizing half only, not how to correct a wrong answer or how an answer is built. It rests on the research paper and on the providers' own crawler and search documentation, read for this guide and linked where used. Anything from our experience on files is written as a tendency, never as a promise.

The short version

Optimizing for a person means publishing accurate facts as self-contained sentences, on pages assistants can reach, with independent sources that agree. Nothing else has stronger support.
The one research paper behind the term tested nine tactics on two engines and reports gains of up to 40 percent in how visible a source is. It did not study personal names, and keyword stuffing did not help.
Google says there is no special optimization or markup for AI Overviews and AI Mode. Ordinary search eligibility is the requirement.
Build a claim sheet first: each fact about you, the source that proves it, and the sentence you want quoted. Then publish from the sheet.
Never use fake profiles, undisclosed paid mentions or text written for machines and hidden from readers. They put your name at risk and no provider page we read endorses them.

Up to 40%the visibility gain the GEO authors report for their best methods on their benchmark

Ninecontent tactics the paper tested, such as adding quotations, statistics or citations

Twogenerative engines tested, so results may not carry to other assistants

Aggarwal et al., "GEO: Generative Engine Optimization", arXiv 2311.09735, version of June 2024 (KDD 2024). The figures describe the authors' benchmark, not individual people.

## Generative engine optimization for a person is evidence work, not a trick An assistant asked who a person is has very little to work with. A company has a catalog, reviews, press and filings. Many individuals have a profile, an old interview and a directory entry. When the evidence is thin, the answer is built from whatever is easiest to extract, and one page can decide the whole paragraph. A clear, accurate, supported page raises the odds that the answer reflects you. Something that looks engineered gains nothing and may become the page an investor screenshots. Three jobs follow, and the rest of this guide is organized around them. - **Say it clearly.** State each fact once in a sentence that makes sense out of context. - **Make it reachable.** Let the right crawlers read the page, and tell search systems which pages are you. - **Make it confirmable.** Point to independent sources that say the same thing, so the claim does not rest on you alone. You will also see "AEO" (answer engine optimization) and "AI SEO" used for the same idea. We use GEO because the research literature does. ## What the research on GEO shows, and what it does not The paper that named the field, "GEO: Generative Engine Optimization", is a paper by Aggarwal and colleagues, posted on arXiv and presented at the KDD 2024 conference. It treats a generative engine as a black box that reads several sources and writes one answer, and it asks which edits to a source page make that page show up more in the answer. The authors built a benchmark of varied queries and tested nine content changes on it. Their headline result is that the best methods raised a visibility measure by up to 40 percent. In the version we read, the strongest single change was adding quotations to a page, with a visibility score rising from 19.3 to 27.2, about 41 percent. Adding statistics and citing sources also ranked among the best methods. Keyword stuffing, the staple of older SEO, did not help, and the authors say it offers little to no improvement. They also report that the effect depended on the domain: quotations helped most on queries about people and society, citations helped most on factual queries, and statistics helped most on law and government queries. Read that for what it is. It measures a source's share of a generated answer, not a person's accuracy or reputation. It tested two generative engines, a setup the authors built and Perplexity.ai, and did not measure search rankings, and the authors list both as limits. The version we read is dated June 2024, so newer engines are untested. And the changes it rewards are changes toward better writing, because real quotations, figures and named sources are things a careful author does anyway. The tactics that worked make a claim easier to trust and easier to lift. None makes a weak or false claim win. That is the line this guide keeps. ## Build a claim sheet before you write a single page A claim sheet is a table of what is true about you, how you can prove it, and the sentence you would be happy to see quoted. It is the document everything else is published from, and it is the step most people skip because it feels like paperwork. It is also what stops two of your pages from contradicting each other. Keep it to the facts a stranger needs, and leave out adjectives. An assistant can lift "She leads product design at an energy software company" and cannot do much with "a passionate visionary". The example below is an invented scenario with generic details: an independent product designer who shares a name with a photographer.

Claim	Independent proof	Where it lives	Quotable sentence
Current role	Client page or company team page	Home and About	Name is an independent product designer working with software companies.
Location	None needed; self-stated	About	Based in Leeds, United Kingdom.
Known for	Conference program, published interview	About and Talks	Spoke on onboarding design at a named conference.
Not to be confused with	Self-stated, links to both profiles	Home and About	Not the photographer of the same name.

A claim with no proof stays on the sheet as a plain statement of your own, never dressed up as an achievement. And every claim has one canonical wording that you copy wherever the fact appears, because inconsistent wording is where assistants improvise.

How to build the claim sheet in one sitting

List every statement a stranger would needWrite down role, organization, location, field, notable work, credentials and the one line that separates you from a namesake. Aim for eight to fifteen items.
Attach a proof to each itemName the independent page that confirms it: a conference listing, a company page, an interview, a professional register. Mark items that rest only on your word.
Write the canonical sentenceOne short, complete sentence per item, with your full name or a clear pronoun, no marketing adjectives, and no claim the proof does not support.
Check it against what already existsSearch the main facts on your existing profiles and press mentions. Where an old page disagrees, decide whether to update it or to explain the change.
Assign each item a homeChoose the page that will carry it and the pages that will repeat it. Facts that change often, such as a job title, get one source page and links from the rest.

## Write pages an assistant can quote without losing the meaning An assistant that retrieves your page reads it in pieces, and a sentence is quoted on its own, away from its heading. If it says "she" and the name is two paragraphs up, the meaning does not survive the cut. The test: paste any sentence into a blank document and ask whether a stranger would know who and what it is about. If not, name the subject. Google's helpful content guidance asks whether it is clear who created a page, and it encourages accurate authorship information such as bylines, with links to background on the author. The page also warns against deceptive authorship, such as fabricated creator profiles, which is a good rule for the whole exercise. - **Open with the answer.** The first paragraph of your About page says who you are, what you do and where. - **Use real quotations and real numbers, with sources.** The research found quotations and statistics helpful, and for a person the honest version is a short quote from a talk, a client, or an article, attributed and linked, and a figure you can document. An invented or inflated number defeats the purpose and invites a correction. - **Name the things.** Organizations, conferences, publications and dates give a system something to match against other pages. "Leading brands" gives it nothing. - **Date what changes.** "Role as of October 2026" tells readers and retrieval which statement is current. - **Keep the important text as text.** Google's own page says to keep important content in textual form. A biography inside an image or a script-only widget is invisible to most retrieval. Do not write for the machine at the cost of the human. Google's guidance says content made mainly to attract search visits does not fit what its systems reward, and a page stitched from prompts reads that way to the investor who clicks through. ## Let the right crawlers in, and decide on training separately None of the writing above matters if the page cannot be read. Access to AI search runs through crawlers, each with its own name and its own robots.txt rule, and the providers document them as separate switches. Your decision about a search crawler and your decision about training are two decisions.

Crawler	What the provider says it does	If you block it	For a personal brand
OAI-SearchBot	Surfaces sites in ChatGPT's search features	Site is not shown in ChatGPT search answers, though it may still appear as a navigational link	Usually allow, if you want to be findable there
GPTBot	Crawls content that may be used to train OpenAI models	Signals the content should not be used for training	Your choice, separate from search
ChatGPT-User	Visits pages when a user asks in ChatGPT	Robots.txt may not apply, and it does not decide Search eligibility	Not the control for search visibility
Claude-SearchBot	Crawls to improve search result quality	May reduce visibility and accuracy in search responses	Usually allow, for the same reason
Claude-User	Accesses a site when a Claude user asks a question	May reduce retrieval for user-directed search	Usually allow
ClaudeBot	Collects content that could contribute to training	Excludes future content from training sets	Your choice, separate from search

The descriptions come from OpenAI's crawler documentation and Anthropic's help page on its crawlers, as read for this guide. OpenAI says a robots.txt change can take about 24 hours to take effect for search, and Anthropic says its bots honor robots.txt. Check the pages before you change anything, since providers add and rename agents. For Google, the rule is stricter and simpler. Its AI features page says a page must be indexed and eligible to appear in Search with a snippet, and that there are no additional requirements to appear in AI Overviews or AI Mode, nor special optimizations necessary. It lists the controls: robots.txt for Googlebot, nosnippet and its variants, and Google-Extended for training and grounding in other Google systems. A site that blocks snippets to look discreet is also hiding itself from the answers it wants to appear in. The common mistake is a global block, set by a hosting default or a privacy plugin, that closes every crawler at once. Decide on purpose: allow search crawlers if you want to be found accurately, and opt out of training separately if you wish. ## Use structured data and identity links to say which pages are you Structured data states facts in code that a machine can read without guessing. Schema.org's Person type defines properties such as jobTitle, worksFor, alumniOf and knowsAbout, and a sameAs property that points to pages that clearly identify the person, such as a verified profile or an entry in an open knowledge base. Be realistic about what this does. Google's AI features page says there is no special schema.org structured data you need to add for AI Overviews or AI Mode, and recommends that any structured data match the visible text of the page. Its ProfilePage documentation describes markup for pages where a single person or organization shares first-hand perspectives, with that person as the main entity, and it says rich results are not promised. So markup is a tidy way to state what the page already says. It is not a lever. It earns its place in the shared-name case, where a page that states in text and markup who you are, and links to your real profiles, helps a system tell you from a namesake. Mark up only what is visible. Link both ways: the About page lists your profiles and each profile links back. Use one name form everywhere. And separate yourself from the namesake in words first, since one sentence naming your field does more than any markup. ## Earn confirmation from independent sources, and refuse the shortcuts A claim that appears only on your own site is a claim made by you. Statements confirmed on more than one independent page tend to be safer to repeat, for retrieval systems and for readers. The most useful pages are ones you did not write: a conference program, a company team page, an interview, a professional register. You can earn these in ordinary ways. Accept speaking invitations and check that the program lists you correctly. Ask a client or employer whether a team page can name your role. Answer journalists and podcast hosts who approach you with the same canonical facts. Check that directories you qualify for carry the right title and location. The shortcuts are where people hurt themselves, and they are tempting because they are marketed as GEO. - **Paid placements presented as editorial.** If the publisher does not disclose the payment, the page is a liability for both of you. Disclosed sponsorship is fine. - **Fake or purchased profiles and reviews.** They breach platform rules and attach to your name when found. - **Mass-produced pages.** Dozens of near-identical articles about yourself add noise, not signal. - **Text aimed at assistants and hidden from readers.** We have not found a provider page that endorses it, it reads as deception once someone views the source, and it is the kind of thing a screenshot travels with. - **Editing Wikipedia about yourself.** Wikipedia's conflict of interest guideline says you should not create or edit articles about yourself, and paid editors must disclose their client. ## Which tactics are worth your time, ranked by evidence, effort and risk The table ranks common GEO tactics on support, effort and risk. The ranking is our judgment from the sources above and from experience on files, not a measured league table.

Tactic	Support	Effort	Main risk
Claim sheet and consistent wording	Strong logic, consistent with provider guidance on clear authorship	Low	Few; stale facts if not maintained
One clear About page with a byline-style bio	Matches Google's helpful content guidance	Low	Vague or inflated wording
Attributed quotations and documented figures	Best result in the GEO paper, on its benchmark	Medium	An unsupported number invites correction
Allowing search crawlers	Documented by OpenAI and Anthropic	Low	Training and search treated as one decision
Independent confirmations (talks, team pages, interviews)	Strong for trust; not measured for names	High, slow	Undisclosed payment
Person and sameAs markup	Weak; Google says none is needed	Low to medium	Markup that does not match visible text
Posting volume	No primary source found	High	Thin pages that dilute the real ones
Hidden or machine-only text	None	Low	Reads as deception, hurts trust

What helps is what a careful professional would do for human readers. What promises a shortcut has no support and a visible downside. If a vendor sells the bottom of that table, ask for the method behind the numbers. ## A worked scenario and a realistic timeline This is an invented scenario with generic parties. An independent product designer has a portfolio of adjectives, a professional profile, and a name shared with a photographer who has more press. Asked who the designer is, an assistant returns a blend, or the photographer. Nothing false is said, so there is no complaint to file: the evidence is thin and shared. The work follows this guide. Build the claim sheet. Rewrite the About page around the canonical sentences, with a line naming the field that separates the designer from the photographer. Add an attributed client quote and one documented figure. Allow search crawlers. Ask a conference organizer to correct a misspelled name. Link profiles and page both ways. Test with the same fixed questions before and after. Experience suggests the order below. These are tendencies, not promises, and they depend on how often each system refreshes its view of your pages.

When	What happens	What to check
Week 1	Claim sheet written; old pages compared with it; crawler settings fixed	Search crawlers are not blocked; no page contradicts the sheet
Weeks 2 to 3	About page rewritten; profiles updated to the canonical wording; markup added to match the page	Page is indexed; profiles link to it and it links back
Weeks 3 to 6	Corrections requested from the third-party pages that carry errors; one or two real confirmations earned	Third-party pages now say what the sheet says
Weeks 6 to 10	Search systems and assistants begin to retrieve the new pages; answers start to shift in some places and not in others	Re-run the fixed questions on the same assistants
Month 3 onward	Answers settle on a clearer description, or the remaining gap shows which fact still lacks an independent source	Add the missing confirmation, update the sheet quarterly

Nothing in the first three weeks depends on any assistant. And the end state is not "the assistant says what I wrote": it is that a clear, supported description exists to retrieve, and the remaining gaps show where to work next. Answers drawn from training alone change on the provider's schedule, which you cannot set. ## How to tell whether it worked You cannot tell from one search, because the same question can return different answers from one run to the next. Use a short, fixed list of questions a stranger would ask, run it the same way before and after the changes have had time to be read, and compare claim by claim. Track three signals in order: does the answer describe the right person, which matters most for a shared name; does it repeat your canonical facts; and which sources does it cite. Your own pages or independent confirmations mean the structure is working, and an old directory entry means that page needs attention. Scoring the questions properly is a separate method, and our AI search reputation work covers it for people who do not want to repeat it by hand. Resist a single "AI visibility score": one built on a vendor's prompts may tell you more about the prompts than about you. When an answer is false rather than thin, more optimization is the wrong response: it calls for a correction through the provider's route or at the source. Optimization is for evidence that is accurate but weak. ## Your next step: write the claim sheet this week You do not need a tool or a budget to start. The first useful hour is a document with ten lines in it.

If you want a second pair of eyes first

Write the claim sheet, then read it as a stranger who knows nothing about you. If it feels thin, or you share a name with someone else, a short private review of what assistants find today will show where the evidence is weakest. Our personal brand work starts from exactly that picture, and Reputation Partners builds the owned page that states the facts.

After the claim sheet: fix crawler access, rewrite the About page, align the profiles, ask one independent page for a correction or confirmation, and re-run your questions in six to eight weeks. ## Questions people ask ### Is generative engine optimization the same as SEO? They overlap, but they are not identical. Search engine optimization aims at ranking pages in a list of results. GEO aims at being used and cited inside a written answer. Google says the SEO basics still apply to its AI features, so the foundation is shared. What changes is that sentences are lifted out of context, so each one must make sense alone. ### Do I need a Wikipedia page to show up in AI answers? No source we read says you do. Some vendor reports claim large effects from Wikipedia, but they publish no method. Wikipedia's conflict of interest guideline also says you should not create or edit articles about yourself. Independent coverage that satisfies those rules is worth earning for its own sake; do not chase a page for AI. ### Should I block AI crawlers to protect my name? Blocking search crawlers can keep your own pages out of AI search answers, which leaves other people's descriptions to fill the space. OpenAI and Anthropic treat training and search crawlers as separate switches, so you can opt out of training and stay findable. Decide each one on purpose, and check the provider's page before changing settings. ### Does structured data make AI assistants use my page? Not on the evidence we read. Google says AI Overviews and AI Mode need no special schema, only that structured data match visible text. Person and sameAs markup is a tidy way to state who you are and link your profiles, which helps with a shared name, but it is a supporting detail, not a lever. ### How long until an assistant describes me differently? It varies by assistant and by whether it searches the web for the answer. Pages that are retrieved live can start to appear within weeks once they are indexed and reachable. Answers that come from training memory change only when the provider updates its models, on a schedule you do not control. Re-test over months, not days.