Credible Roots / Answer Engine Optimization
Answer engine optimization for companies
The AEO agency that publishes its own method
Most answer engine optimization is a content quota with a new name on it. Ours is a technical floor, an entity the models can resolve, and pages built on data your competitors do not have — with the sources named, the assumptions printed, and the corrections logged in public where anyone can check them.
One client engagement, named and published with permission. All four figures are checkable on the live site right now. The full case study, including what that correction cost.
Answer engine optimization is the work of becoming a source that ChatGPT, Perplexity, Google AI Overviews and Gemini retrieve and cite when someone asks a question in your category. It is not a ranking exercise, because answer engines do not rank pages: they retrieve passages at the moment of asking, compose a reply, and attribute a handful of sources. Three things decide whether you are one of them. Access, meaning the crawlers behind each engine — GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and Google-Extended — can actually fetch your pages, which is a separate decision from Googlebot and is frequently blocked by a host setting nobody chose. Extractability, meaning a passage near the top of the page states a complete, specific, checkable fact in full sentences, because names, dates and numbers survive extraction and adjectives do not. And corroboration, meaning the same claim appears the same way somewhere the engine already trusts, which is why published research repeatedly finds citations skewing toward earned third-party coverage rather than a brand’s own domain. The uncomfortable part, which we would rather you hear from us: AI referral traffic is still around 1% of total web traffic. AEO is a positioning investment in who gets named as the authority, not a near-term traffic channel, and any agency selling it as the latter is selling you something else.
- Retrieval, not ranking. The unit is a quotable passage, not a page position.
- Crawler access is decided separately from Googlebot, and is the most common silent failure.
- Corroboration beats on-site optimization. The research is consistent on this.
- It is not a traffic channel yet. Buy it for position, not for sessions this quarter.
On this page
- What answer engine optimization actually is
- AEO, GEO and SEO: one floor, three jobs
- What we build
- A worked example: heatpumpmath.com
- How to choose an AEO agency
- What the reporting should look like
- What the research actually shows
- What nobody can promise
- Who this is for, and who it is not
- Questions companies actually ask
What answer engine optimization actually is
A search engine returns a list and lets the user choose. An answer engine composes a reply and chooses for them, citing a few sources as it goes. That single difference changes what the work is.
Optimizing for a list means competing for a position against a query. Optimizing for an answer means being retrievable, quotable and corroborated at the moment the question is asked. The page is no longer the unit. The passage is. A page that ranks fourth for a query can still be the source an answer engine quotes, and a page that ranks first can be skipped entirely because nothing in it survives extraction.
That is the whole discipline, and most of what is sold as AEO is a blog quota with the acronym stapled on.
AEO, GEO and SEO: one floor, three jobs
The three share a technical foundation and diverge above it. It is worth being explicit about which you are buying, because they fail differently and they are measured differently.
| SEO | AEO | GEO | |
|---|---|---|---|
| The unit | A page, ranked against a query | A passage, retrieved and quoted | An entity, described consistently enough to be generated about |
| What wins | Best match for the intent behind the query | A liftable, specific, corroborated claim | Being the consensus answer across the sources a model was trained or grounded on |
| Access controlled by | robots.txt and rendering | Separate AI crawler directives, set independently | Both, plus whatever third parties publish about you |
| Corroboration | Helps | Close to decisive | Decisive |
| Measured by | Impressions, position, clicks | Who gets named, and which sources the answer cites | Share of prompts where you appear at all |
| Honest timeline | Months | Weeks for access faults, months for position | Slowest of the three |
In practice the terms are used loosely and often interchangeably, including by people selling them. We run all three as one engagement because the floor is shared: a site the engines cannot fetch, render or resolve fails at all three simultaneously.
What we build
Five things, in this order, because each is wasted without the one before it.
The access audit
Whether GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot, Google-Extended and Applebot-Extended can fetch your pages, and whether those pages exist without JavaScript. This is the most common silent failure we find, and it is usually a CDN bot-management default or a security product nobody configured deliberately. Which of those you should allow is a decision with real trade-offs, not a checkbox.
Entity resolution
Structured data describing a real organization and real people rather than a generic web page, one canonical identity, consistent details everywhere you appear, and a Wikidata record where the company genuinely qualifies. Models answer about entities. A company that cannot be resolved to one competes only on strings, and gets confused with the three other firms sharing its name.
Extractable writing
Restructuring the pages that matter so each opens with a complete, specific answer in full sentences, with the names, dates and numbers that survive extraction. This is an editorial discipline more than a technical one, and it is the part most agencies skip because it means rewriting rather than publishing.
Original data, where there is data to be had
The most defensible asset in any category is a number nobody else holds. Sometimes that is a public API nobody in the sector has pulled. Sometimes it is measuring something yourself. It is more work than another opinion piece and it is the only part a competitor cannot replicate the week after they notice you.
Corroboration, and a baseline you can re-run
Getting independent sources to state the same facts the same way, because that is what the research says moves citation. Then a recorded baseline of the prompts your buyers actually ask, re-run on a fixed schedule so change is visible rather than asserted.
A worked example: heatpumpmath.com
Named client, published with permission
heatpumpmath.com answers one question for American homeowners: does a heat pump cost less to run than the system already in the house. Every comparable tool we found ran on national average energy prices, which is why they disagree with each other.
Residential electricity and gas prices come from the U.S. Energy Information Administration’s Open Data API v2, pulled by a script in the project, covering 51 states including the District of Columbia. The heating-season window is fixed at December to February for a published reason: residential gas volume collapses in summer, fixed customer charges then dominate the per-unit price, and using the latest month would bias every comparison toward heat pumps. Four assumptions are printed with their values. No winner is declared inside a 10% margin. Six states are excluded from gas comparisons entirely. The computed figures are released under CC BY 4.0.
On 31 August 2026 we published a correction: the pipeline had been using the most recent monthly figures, which for gas fell in May and June. Fixing it cut the number of states where a heat pump wins on running cost under cold conditions from nineteen to four — making the site’s own headline finding substantially weaker. It went up with both numbers side by side.
We lead with this rather than a dashboard because it is the part an agency cannot fake in a pitch. Anyone can show a screenshot of a chatbot saying their client’s name. Very few can point at a live methodology page, a live corrections log and a data pipeline you can go and inspect — which is, not coincidentally, exactly the shape of thing answer engines quote.
How to choose an AEO agency
Six questions worth asking anyone selling this, including us. They are cheap to ask and they separate the field quickly.
- Show me the crawler audit you would run, before we sign. If the first deliverable is a content calendar rather than an access check, they have skipped the only step that can make everything else worthless.
- What will you report, and what will you refuse to report? An agency that promises a ranking number for an answer engine either does not understand the systems or is counting on you not checking.
- What have you been wrong about, and where is that written down? Anyone publishing for a year has been wrong. No corrections means either not enough published or nothing corrected.
- Which numbers in this pitch are yours, and which came from another agency’s blog? This industry recycles statistics until the original is unfindable. The 7-11-4 rule is the clearest worked example.
- What in this plan could a competitor copy in a week? Headings, FAQ blocks and schema are table stakes. If the whole plan is copyable, it will be copied.
- What share of my result will come from things on my own site? The honest answer is: less than you would like. Anyone claiming otherwise has not read the research below.
What the reporting should look like
This is where most AEO engagements quietly fail, because the honest metrics are uncomfortable and the flattering ones are meaningless.
| What we report | What we will not report |
|---|---|
| Share of a fixed prompt set where you are named at all | A “ranking” inside ChatGPT, which does not exist |
| Which sources the answers cite instead of you, by name | A single screenshot of a good answer, presented as a trend |
| Crawler access status per engine, with the date checked | An access claim inferred from robots.txt alone |
| Which pages changed, and what the passage looked like before | Word counts and pages published |
| Independent mentions gained, with links | “Brand sentiment” scores with no stated method |
The prompt set is fixed at the start and re-run unchanged, because a prompt set edited between reports can be made to show anything. These systems are probabilistic and the same prompt returns different answers on different days, so the honest form of this is directional evidence over time, not a number that goes up every month.
What the research actually shows
Published research
| Finding | Figure | Source |
|---|---|---|
| AI citations coming from earned media rather than owned content | about 84% | Muck Rack, May 2026 |
| Brands more likely to be cited through third parties than their own domain | 6.5x | AirOps, October 2025 |
| ChatGPT citation presence in professional services prompts | under 4% | Similarweb, May 2026 |
| AI referral traffic as a share of total web traffic | about 1% | Search Engine Land, 3.3bn sessions |
Read those together and the conclusion is awkward for a page selling this service. Most of the citation is earned elsewhere rather than won on your own site, and the traffic it currently sends is small. We would rather you buy this understanding that, because the agencies quoting you a traffic forecast are the ones you should worry about.
What nobody can promise
- That a given model will name you. These are probabilistic systems that change without notice and answer the same prompt differently on different days.
- A ranking position in an answer engine. There is no ranking to hold.
- A traffic number. See the table above. Anyone forecasting AI referral traffic for your category is guessing with confidence.
- Removal of a competitor from an answer. Not a lever anyone has.
- A fixed date. Access faults resolve on the crawler’s schedule. Position does not.
Who this is for, and who it is not
| A good fit | Not yet | |
|---|---|---|
| Category | Buyers research before they buy, and ask questions a model can answer | Transactional or purely local demand, where nobody asks an engine anything |
| Material | Has data, measurements or operating detail nobody has published | Only approved marketing language, signed off by committee |
| Site | Willing to change the technical floor and rewrite existing pages | Wants new content only, existing pages untouched |
| Earned coverage | Has some, or will invest in getting it | Expects on-site work alone to be enough |
| Horizon | Buying position over the next year | Needs attributable pipeline this quarter |
| Evidence | Comfortable with directional measurement, honestly reported | Wants a number that rises every month |
Some calls end with us saying this is the wrong purchase. If you need demand this quarter, buy demand generation; it works and this does not, on that timescale.
Questions companies actually ask
What is the difference between AEO, GEO and SEO?
SEO competes for a position in a ranked list. AEO makes a passage retrievable and quotable so an answer engine cites it. GEO is about the entity itself being described consistently enough across the web that models generate accurate answers about you. They share one technical floor — crawlable, renderable, resolvable — and diverge above it. The table above sets out how each is measured and how each fails.
Is answer engine optimization actually worth buying yet?
It depends what you are buying it for. AI referral traffic is around 1% of total web traffic, so if the goal is sessions this quarter the honest answer is no. If the goal is being the name that comes back when a buyer asks an engine who the credible option is, the work compounds and the category is far less contested than organic search. We would rather lose the sale than have you buy it on a traffic forecast.
Can you guarantee ChatGPT will mention us?
No, and neither can anyone else. Answers vary run to run, models change without notice, and no vendor has a lever that inserts a name. What can be committed to is the access, the structure and the corroboration that make citation possible, plus an honest baseline that shows whether it is happening.
How is this different from just doing good SEO?
A large part of it is good SEO, and we say so. The differences that matter are crawler access decided separately from Googlebot, writing restructured for passage extraction rather than page ranking, entity resolution treated as a deliverable rather than a nice-to-have, and measurement against a prompt set instead of a rank tracker.
Do we need original data to compete?
No, but it is the strongest version and the only part competitors cannot copy quickly. Where there is no dataset to gather, the differentiator becomes operating specificity: the details someone who has actually run the thing knows and a generalist writer does not. What does not work is a well-formatted restatement of what already ranks.
What does an engagement cost?
$2,000 a month, with a three month minimum, then month to month. One price covering everything on this page — there are no tiers, no setup fee and no per-article rate, and the first call is free. None of the five agencies we assessed in this field publishes any price at all, which is why ours is on a page of its own, along with what it does not buy.
Read the method before you buy it
- The heatpumpmath.com case study, in full
- How answer engines pick sources, in detail
- Which AI crawlers should you allow
- Does llms.txt actually do anything
- The search half of the same engagement
- A worked example of a statistic nobody can source
- Generative engine optimization: the entity half
- The metrics, with formulas and sample sizes
- How to choose between AEO agencies, us included
- The statistics, traced to source
- The legal vertical, and why the playbook changes
Sources
- OpenAI. Overview of OpenAI crawlers: GPTBot, OAI-SearchBot and ChatGPT-User.
- Anthropic. Crawler documentation and how site owners can block it.
- Google. Overview of Google crawlers and user-triggered fetchers, including Google-Extended.
- Google. Search Essentials.
- heatpumpmath.com. Methodology: data sources, constants, assumptions and where the model fails.
- heatpumpmath.com. Corrections log.
- U.S. Energy Information Administration. Open Data API v2.
How this was made: written from the work we do for clients, with every figure in the proof strip checkable on the client site it describes. The research figures are attributed to their publishers and linked on the AI search visibility page. Drafted with AI assistance, then reviewed and approved by the named author before publication. Our editorial standards.
Start with the audit, not the pitch
The first call is a diagnostic. We will check whether the engines can reach you at all, run the prompts your buyers ask, show you who gets named instead, and tell you honestly whether this is worth buying for your category yet.
Book a 30 minute call