Generative engine optimization for brands

Generative engine optimization for a brand is the work of making your company the easy, accurate answer when an AI assistant is asked a question in your category. It is mostly content and evidence work: pages that state commercial facts in sentences that stand alone, independent sources that confirm them, and nothing on the site that stops a crawler from reading either. It is not a way to make an assistant say something untrue.

The central difference from ordinary brand search is that most of the questions that matter do not contain your name. A buyer asks for the best option for a given need, a comparison of two vendors, or an alternative to the market leader. The assistant answers from a handful of pages, and whether you are in that handful depends on what exists about your category, not on how well you rank for your own name.

This guide covers the optimizing half for companies, not auditing the answers you have or how an assistant builds a reply, and not a single person's name. It rests on the original research paper and the providers' documentation, read for this guide and linked where used. Our own experience is written as tendency, never promise.

Ninecontent tactics the GEO paper tested against an unoptimized baseline

17.7 vs 19.3keyword stuffing scored below the baseline on the paper's main visibility measure

Noneadditional requirements Google lists for appearing in AI Overviews or AI Mode

Sources: Aggarwal et al., "GEO: Generative Engine Optimization", arXiv preprint 2311.09735, version of June 2024 (scores are the paper's position-adjusted word count); Google Search Central page on AI features. The scores describe the authors' benchmark, not your company.

Generative engine optimization for a brand is won on questions that do not name you

A brand is asked about in six different ways, and only one of them contains the brand name. The rest are category questions, and an assistant answering them has to choose which companies to mention. That choice is made from the pages it can find on the topic, which is why a perfect homepage and a weak category presence often produce an assistant that knows you exist but never volunteers you.

Related readAI Answers About Job Candidates: Audit and Fix

Sort your questions before you touch a page. For each type below, the table shows who usually owns the evidence, which page answers it and where the fix sits.

Two of these rows sit mostly outside your own domain. If you only work on the pages you control, you cover the use-case, comparison and direct-fact rows and leave the category, alternative and trust rows to whoever else writes about you. The rest of this guide is ordered by that division: first the pages you own, then the proof you earn.

What the research shows for brands, and what it does not

The term comes from a 2023 preprint by Aggarwal and colleagues, which proposed generative engine optimization and a benchmark of 10,000 queries across 25 domains. It tested nine tactics against a baseline of unmodified content. The authors report that their best methods improve visibility by up to 40 percent. Read in detail, the paper's table shows the best methods improving a position-adjusted word count by 41 percent and a subjective impression score by 28 percent over the baseline, on the authors' own measures.

Three results matter to a company.

• Evidence-style additions helped. Quotation addition scored 27.2 against a baseline of 19.3 on the paper's main measure. Adding statistics scored 25.2 and citing sources scored 24.6.
• Keyword stuffing hurt. It scored 17.7, below the baseline. On the Perplexity experiment, the paper reports keyword stuffing about 10 percent worse than baseline.
• Smaller sites gained most. The paper reports that citing sources raised visibility for sites ranked fifth in the results by 115.1 percent, while the top-ranked site lost 30.3 percent on average. A challenger brand has more to gain from clear sourcing than an incumbent.

The paper also says the best tactic depends on the domain: quotations did best for people-and-society queries, statistics for law and government. Test in your own category.

The limits are plain. It is a preprint, not a peer-reviewed publication, according to its listing. It tested the authors' own engine, built on a 2023 model and the top five Google results, plus Perplexity. It measured how much of an answer drew on a page, not whether a buyer chose the company.

Related readAI answers about doctors and lawyers: what to fix

Build the pages that answer category questions, in this order

Start with the page that states what you are. A category definition page says, in the first two sentences, what the company does, for whom, and in which category a buyer would look for it. It names the category in the words buyers use, not the words the marketing team prefers. An assistant asked to list companies in a category cannot list a company whose own site never says which category it belongs to.

Next, build use-case pages, each handling one situation with named constraints: customer size, task, integration, regulation. A "Solutions" page with six use cases in one paragraph each is rarely the best source for any of them.

Then add the comparison page, which is the one most brands do badly. A comparison that is plainly one-sided is easy to discount, by a reader and by a system that is also reading rivals' pages and independent reviews. A useful one states the criteria first, gives the facts that can be verified for each side, says where a rival is the better fit, and carries a date. Do not name a rival in a way you cannot support, and do not state a rival's price or feature unless you checked it that month. Where you are the better fit, say for whom and why, with the evidence.

Last, a pricing and specification page in plain current text. A brand that hides price behind a form asks assistants to fill the gap from an old forum post.

1. Answer firstPut the direct answer to the page's question in the first two sentences, with the company named in the sentence and not only in the header.
2. State the constraintName who the statement is true for: the customer size, the region, the plan, the version. A claim with no condition is the first to be wrong.
3. Attach the proofGive a figure with its source and date, a quotation with its speaker's role, or a link to the standard or document behind a claim.
4. Date itShow when the page was last reviewed. Assistants and readers both need to know whether a price or feature is current.
5. Keep it in textKeep the key facts in the page text, not only in images, scripts, tabs that need a click or a PDF.

The proof step is where research and ethics meet. A customer quotation must be real and used with permission, and a statistic must come from a source you can name.

Related readWrong company information in ChatGPT: how to fix

Cover the sub-questions, because assistants search in several directions

Google's page on AI features says AI Overviews and AI Mode may use what it calls query fan-out, which it describes as issuing multiple related searches across subtopics and data sources. The practical reading for a brand is that one buyer question becomes several searches, and your page can be a source for one of them without being the best result for the main phrase. Other providers describe their systems differently, and none publishes the full mechanism, so treat this as a reason to be thorough and not as a recipe.

Build a sub-question map for each category question: list what someone would need to know to answer it well, such as price, integrations, security evidence, migration and who it is wrong for. Each is a short, complete page or a clearly headed section.

The table uses an invented product type. Each cell in the third column names a specific page, and a gap there is a content task. "Who it is wrong for" is the line most brands refuse to write and most buyers trust most.

Treat commercial facts as a register with an owner and a review date

A person has a short list of facts that rarely change. A company has dozens that change all the time: prices, plans, feature names, integrations, office addresses, leadership, certifications, retired products. An assistant that quotes a price from an old page is repeating a fact that was once true, and the damage is the same as a false statement.

Keep a fact register. One row per fact a buyer might ask about, with the exact wording you want used, the page that holds the canonical version, the other places the fact appears, the person who owns it and the date it was last reviewed. When a fact changes, the register tells you every place to update. It also answers the question that a review exposes: which of our pages disagrees with another one.

Related readGoogle AI Mode reputation: audit what it says

Three situations deserve a rule.

• Retired products and old prices. Do not delete the page and leave the old facts in circulation elsewhere. Keep the page, state clearly that the product or price is no longer offered and give the replacement, so a crawler that reads the old URL reads the correction.
• Naming. Pick one legal name and one product name and use them everywhere: site, listings, press boilerplate, app store entries, partner pages. Variants are how an assistant ends up describing two companies, or a company and its former name, as separate things.
• Numbers in public claims. Count of customers, years in business, awards. Each should carry a date or a source. A figure that has been rounded up for three years is a liability that a trust question will surface.

Earn the proof you cannot write yourself

An assistant can read your site and then look for a second voice. For a company, the second voice is richer than for a person: review platforms, trade publications, directories, partner and integration pages, standards bodies, customer case studies on other domains, and community discussions. In category and alternative questions it is often the main evidence.

Work through them by the lawful routes only.

• Trade and industry press. Offer genuine data, a named expert, a launch with something checkable. A press release written to look like news does not change what a publication will print. Disclose paid placements as paid.
• Directories and listings. Claim your profile where one exists, fill it in with the register's wording and keep it current. Wrong category, old description or duplicate entries are common and entirely fixable.
• Partner and integration pages. If you integrate with another product, ask for an accurate listing on its side. These pages are specific and hard to fake.
• Customer evidence. Real case studies with the customer's agreement, and an honest, steady request process for reviews on the platforms your buyers use. Ask every customer, not only the happy ones, and follow each platform's rules on how you may ask.
• Reference sources. Wikipedia's conflict-of-interest guideline very strongly discourages owners, employees and contractors from editing articles about their own company, and its terms of use require paid contributions to disclose the employer or client. The route it describes is a disclosed suggestion on the article's talk page, which other editors may decline. An undisclosed edit risks reversal and a story about the attempt.

Make the commercial site readable by the systems that feed assistants

Brand sites fail on access in ways a personal site does not: pricing in a script, documents in PDFs, content behind a cookie wall or login. A crawler that cannot read a page cannot cite it, and the assistant fills the hole with a third party.

Check four things, briefly, since the mechanics are covered elsewhere.

First, eligibility. Google says a page must be indexed and eligible to be shown in Search with a snippet to be a supporting link, and that no additional requirements or special optimizations are needed. It also says structured data should match the visible page text, and that there is no special schema.org markup needed for these features. It recommends keeping important content in text form and, for businesses, keeping Merchant Center and Business Profile data current. Those are the commercial data sources worth auditing.

Second, the crawlers for each assistant. OpenAI documents separate agents: OAI-SearchBot, which surfaces sites in ChatGPT search features, GPTBot, which may be used for training, and ChatGPT-User, for actions a person starts in the product. Its page says the settings are independent, so a brand can appear in search results while opting out of training, and that a robots.txt change can take about 24 hours to take effect for search. Other providers publish their own pages, and policies change, so check the current ones before deciding. OpenAI says a site that opts out of OAI-SearchBot will not appear in ChatGPT search answers, so a blanket block is worth checking first when a brand is missing.

Related readWhat Siri, Alexa and Google Assistant say about you

Third, content in text: move key prices, specifications and policies out of images, tabs and downloads. Fourth, expect no separate AI segment: Google says AI feature traffic is reported in the Performance report under the Web search type, mixed with other traffic.

Which tactics are worth the effort, and which to refuse

Start with the cheap, low-risk work and skip anything that depends on deceiving a platform or a reader.

Four more things to refuse, whatever a vendor promises. Hidden text or instructions aimed at an assistant deceive the system and look deceptive when found. Mass-produced pages written only to cover phrases add no information, and Google's spam policies list generating many pages without adding value, including with generative AI, as scaled content abuse. Its hidden text rule covers content placed solely to manipulate search engines, while ordinary tabs and accordions are not a violation. Google's separate page on generative AI content could not be read for this guide, so check it before publishing AI-assisted pages. Fake threads or sock-puppet community accounts can be exposed, and communities do not forgive them. Attacks on a rival's pages or reviews are the opposite of this work.

A worked scenario and a realistic twelve weeks

This is an invented scenario with generic parties, to show the order of operations. A twelve-person software company sells invoicing to small agencies. Assume 20 buyer prompts, run three times each on one assistant, which makes 60 answers per round. Assume that in the baseline round the company is named in 6 of the 60 answers, mostly for direct fact questions, and that three older facts appear: a retired plan, an old price and a former product name. These numbers are assumptions chosen to show the method, not a forecast and not a benchmark.

Related readWhat Microsoft Copilot says about you: check and correct

In this scenario, suppose the second round at week 6 shows the company named in 12 of 60 answers, and the round at week 12 shows 21 of 60. Again, these are invented to show how to chart your own re-tests.

Baseline6 answers

Week 612 answers

Week 1221 answers

Illustrative numbers from the stated assumptions, not client data and not a forecast. Bar widths are each value divided by 21.

One caution on reading your own version: 60 answers is a small sample and assistants vary from day to day, so a change of a few answers is noise, and a count says nothing about whether the description was right.

Measure the right things, and know where this work stops

Measure four things on a fixed prompt set, repeated at intervals. The share of category and comparison answers that name you. How you are described when named, checked claim by claim against your register. Which pages are cited. And whether the wrong facts have stopped. Add the referral traffic you can see, knowing from Google's own documentation that AI feature traffic is mixed into the Web search type. A monthly re-run is usually enough, and an ongoing program is the kind of work that reputation monitoring is built around.

Accurate criticism stays in the answer. You can respond where it appears and correct what is outdated or wrong, and a false statement is a separate correction job. No one controls which assistant mentions you, so plan for movement.

Your next step: write the register and the six questions

Start with two documents that take an afternoon: the fact register for your twenty most-asked facts, and the six-row question table above, filled in with the real wording buyers use. Between them they show which pages you need and which questions need outside proof.

Related readChatGPT wrong information about my product: how to fix

Questions people ask

Is generative engine optimization different from SEO for a brand?

They overlap heavily. Google says existing search fundamentals remain worthwhile for its AI features and that no extra requirements exist. The difference is the unit: SEO targets a ranking for a phrase, while GEO targets being a source for an answer that may combine several searches and several domains, including ones you do not own.

Can a small brand compete with a market leader in AI answers?

Sometimes. The GEO paper found that lower-ranked sites gained more from adding sources than top-ranked ones, although it tested a narrow set of engines. A challenger usually wins on specificity: pages and independent proof for a narrow use case, rather than a general claim to be the best overall.

Should we publish an llms.txt file or special AI markup?

Google says you do not need new machine-readable files, AI text files or markup to appear in its AI features. We did not review other providers' positions on such files for this guide. Spend the time on clear pages and consistent facts before any extra file, and treat any vendor claim about such a file as unproven until a provider documents it.

How often should we update comparison and pricing pages?

Whenever a fact changes, and on a fixed review otherwise. Show the review date on the page so readers and crawlers can judge freshness. A monthly check of every competitor fact you state is realistic for most teams, and removing a claim you cannot verify is better than leaving it to age.

Do we need to allow every AI crawler?

No. OpenAI documents its search, training and user-initiated agents as independent settings, so a brand can allow search while opting out of training. Decide per crawler, record the reason, and recheck as providers update their pages, since a change can take about a day to apply.

Is it acceptable to ask customers for reviews?

Yes, if you ask honestly and follow each platform's rules. The FTC rule bars paying for reviews that express a set sentiment, fake and undisclosed insider reviews, and threats against negative ones. Ask all customers, never script the opinion, and disclose any incentive where the rules require it. When in doubt, read the platform's current policy and take your own counsel's advice before running a program.