How do AI assistants decide which firms to name?
With web search on, an assistant can run searches of its own, read a few pages it finds and answer from them, often citing those pages. The pages it reads can shape which firms it names, alongside what it learned earlier, so a clear answer to the exact question, on your site or another, gives it something to use.
Key takeaways
- With search on, an assistant can answer from pages it finds and reads, and often shows them as sources.
- Google says its AI features may run several related searches to answer one question.
- In a published benchmark, citing sources, adding quotations and adding statistics made pages more visible in AI answers; keyword stuffing made little or no difference.
- A confident but outdated description of your firm can mislead a prospect; keep the same correct description everywhere.
- Nobody can guarantee a recommendation; build pages that answer real questions and check what the assistants say.
Where this comes from
Public documentation from Google and OpenAI and one peer-reviewed study, each linked in the sources below. Examples marked Illustrative are not measured data.
Make the examples about my firm
Changes the wording of examples only. Stored in this browser only.
What happens between the question and the answer?
With web search switched on, the assistant can search the web itself, sometimes more than once, then read a few of the pages it finds and write one answer from them.
Google describes its AI Overviews and AI Mode as able to run several related searches across subtopics to build one response, an approach it calls query fan-out.1 Those searches are the assistant’s, not your prospect’s, so the question a prospect types and the pages the assistant reads can look quite different.
One question, traced from search to answer.
Illustrative example, not measured data A made-up question, made-up searches and made-up pages, to show the steps.
- 1. The question
“Who’s a good advisor for a recently widowed teacher in your town?”
- 2. Searches the assistant might run
- widow financial planning your town
- fee-only advisor for teachers near your town
- what to do with finances after a spouse dies
- 3. Pages it might read
- A planning firm’s page titled “Financial planning after losing a spouse”
- A local directory listing of advisors in your town
- A news article on teachers’ pension choices
- 4. One answer
“A planning firm in your town works with widowed clients and explains its approach on a page about planning after losing a spouse.”
The pages it read can shape which firms the answer names.
Do all the assistants read the same pages?
Not necessarily. Assistants use different search systems and crawlers, so the same question can bring back different pages, and different firms, from one assistant to the next and from one day to the next.
OpenAI, for example, runs a separate crawler, OAI-SearchBot, for ChatGPT’s search answers, and says sites that block it will not appear there, though they can still appear as plain navigational links.2 A firm that is visible to one assistant can be invisible to another.
What a person sees, and what a crawler can read.
Illustrative example, not measured data A made-up page with five kinds of content.
- Heading and paragraph textReadable: plain text on the page.
- A chart whose numbers appear only inside the imageHard to read: the numbers are pixels, not text.
- An answer inside a PDF behind a sign-up formOften not reached: the form and the download stand in the way.
- Structured data that matches the visible textReadable: it helps a machine read your name, town and services.
- A page on a site whose robots.txt blocks that crawlerNot read by that crawler: the site has told it to stay out.
What makes a page easier to cite?
A page that answers the question in plain words near the top, and shows where its facts come from.
In “GEO: Generative Engine Optimization” (Aggarwal et al., KDD 2024), adding cited sources, quotations and statistics improved a page’s visibility in the authors’ benchmark by roughly 30 to 40 percent on one measure (Position-Adjusted Word Count, roughly how much of the AI answer drew on that page); keyword stuffing did little or nothing.3 That was a research benchmark across many topics, not a test of advisor websites, so treat it as a direction rather than a promise.
The same firm’s knowledge, arranged two ways.
Illustrative example, not measured data A generic mock of your firm’s resources page; the sample answer is placeholder text, not advice.
Built for browsing
An assistant asked “Can I retire at 60 with a pension and $1.2M saved?” finds no page that answers it. The answer may be on page 6 of the PDF.
Built for the question
- Can I retire at 60 with a pension and $1.2M saved?It often depends on three numbers: your pension, your spending and when you claim Social Security.
- Should I take my pension as a lump sum?Placeholder: the answer in one line, then the detail.
- How much should I keep in cash in retirement?Placeholder: the answer in one line, then the detail.
- When should I claim Social Security?Placeholder: the answer in one line, then the detail.
The assistant finds the question as a heading and the answer in its first line.
See the data
| Layout | Can an assistant find the answer? |
|---|---|
| By date and format | Harder: the answer sits inside a PDF, behind a download |
| By question | Yes: the question is a heading, the answer its first line |
Does it matter how an assistant describes your firm?
Yes. A prospect who checks your name with an assistant reads whatever it says, and a confident but outdated description can mislead them in a way silence doesn’t.
An assistant can combine what it reads today with what it absorbed earlier and what other sites say, so an old description elsewhere on the web can outlast the one on your own site. The same short, correct description on your site, your Google Business Profile, LinkedIn and your listings gives it one version to repeat.
One description, four places it might appear.
Illustrative example, not measured data Type your firm’s one-line description to see it in each place. Nothing you type is sent or saved.
- Your website’s About line
Your firm is a fee-only planning firm in your town for teachers nearing retirement.
- A directory listing
Your firm · a fee-only planning firm in your town for teachers nearing retirement
- A professional profile
About: a fee-only planning firm in your town for teachers nearing retirement
- An assistant’s sentence
“Your firm is a fee-only planning firm in your town for teachers nearing retirement.”
What can a firm actually do about it?
Three layers, starting with the one you control most: your own site answers the real questions; your description is the same everywhere; and you show up beyond your own site.
Tick what your firm has today.
Nothing you tick leaves this page. Illustrative: coverage of the layers in this article, not a score and not a forecast.
0 of 3 layers started
Begin with one page that answers one real question.
| Ticked | Own site | Same everywhere | Other sources | Line shown |
|---|---|---|---|---|
| Client-situation pages; same description everywhere | 1 of 3 | 1 of 1 | 0 of 5 | Strong foundation. Now get other sources saying it. |
Reviews and testimonials can count, but the SEC Marketing Rule allows them only under conditions, including clear and prominent disclosures (whether the person is a current client, whether they were compensated, and material conflicts), further disclosure of compensation terms and conflicts, oversight by your firm, and, for promoters paid more than a de minimis amount, a written agreement and a disqualification check. State-registered firms should also check their state’s rules. This is a summary, not the full rule; your CCO decides whether and how the program runs.4
Take this with you
Two prompts to test your own firm, and your layers if you ticked any. Optional; it stays in your browser unless you send it.
Discovery
I live in [town] and I’m [situation]. Who are some independent financial advisors near me who handle [service]? Name specific firms and tell me which sources you relied on.
Identity
What is [firm name]? Who runs it, where is it based and what does it do?
Paste each into ChatGPT, Gemini, Perplexity and Claude. Note who gets named, which sources get cited and whether they describe you correctly. If you’re not named, treat it as a starting reading, not a verdict.
Sending opens your own email app; nothing passes through us. What you type stays in your browser.
Questions principals also ask
Can I pay to be recommended by ChatGPT or Gemini?
We know of no documented way to pay for a place inside an assistant's answer. What you can do is publish clear answers and earn mentions from sources the assistants read.
Does schema markup or an llms.txt file get my firm recommended?
Google says no special structured data or AI text file is needed for its AI features. Markup helps machines read your name, town and services correctly and is worth doing well.
Why does ChatGPT describe my firm wrongly even though my website is correct?
An assistant can combine what it fetches today with what it absorbed earlier and what other sites say. Making the correct description consistent across many places is the practical step; whether it works has to be re-checked over time.
Are client reviews on my Google Business Profile allowed?
They can be, when the program is designed for it. The SEC Marketing Rule allows testimonials in an adviser's advertising only under conditions, including clear and prominent disclosures about client status, compensation and material conflicts, further disclosure of compensation terms and conflicts, oversight by your firm and, for promoters paid more than a de minimis amount, a written agreement and a disqualification check. That is a summary, not the full rule, and state-registered firms should check their state's rules. Your CCO decides; we are not giving legal advice.
How do I know where my firm stands today?
Ask the questions your prospects ask, from the towns you serve, across all the major assistants, and save the answers and citations.
Sources
- Google Search Central, “AI features and your website”, last updated 10 Dec 2025, accessed and checked 7 Oct 2026.
- OpenAI, crawler documentation (OAI-SearchBot), accessed and checked 7 Oct 2026.
- Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, Ameet Deshpande, “GEO: Generative Engine Optimization”, KDD 2024 (arXiv v3, 28 Jun 2024), accessed and checked 7 Oct 2026.
- SEC Marketing Rule, 17 CFR 275.206(4)-1(b), testimonials and endorsements: disclosures, oversight and disqualification; last amended 87 FR 22447, 15 Apr 2022. Accessed 7 Oct 2026.