Building the evidence AI engines will cite

First, the boring prerequisites

Check technical access before investing in a page an engine cannot read. A demonstrated block is actionable; an absent citation alone does not tell you whether the problem is crawling, relevance, source selection or something else.

Use this checklist to find concrete access problems. The items cover different controls, so inspect each rather than assuming one pass proves the next.

The retrievability checklist

1. The page is indexable. No stray noindex, no accidental robots.txt block, no canonical tag pointing somewhere else by mistake.
2. The right crawlers are allowed by name. If your site blocked AI user agents at any point, confirm you blocked the training crawler and not the search crawler. GPTBot and OAI-SearchBot are different decisions.
3. The content is in the raw HTML response, not injected by client-side JavaScript after load. Check with curl -s https://yoursite.com/pricing | grep -i "per month", not with the browser inspector, which shows you the page after JavaScript has run.
4. Nothing stands between the crawler and the text. No consent wall, no interstitial, no soft paywall, no login.
5. Snippet controls are deliberate. nosnippet, data-nosnippet, and max-snippet all restrict what Google's AI features may quote from you. If someone added them years ago to protect featured snippet real estate, that decision now has a larger cost.
6. The page is reachable by an internal link, not only through search. Orphan pages get crawled unreliably.
7. Titles, headings, visible text, and structured data agree with each other. Contradictions here reduce confidence in all of them.

Check your robots file deliberately. OAI-SearchBot controls eligibility for ChatGPT search; GPTBot controls crawling for potential training use. A block can explain an access problem, but our data does not establish it as the most common cause of missing brand mentions.

Write sections that survive being read alone

The unit of work is a section with a heading. It has to meet one standard: somebody reading only that section, with no page title, no navigation, and no surrounding paragraphs, still gets a correct and usable answer.

The same information, written both ways:

Reads poorly in isolation	Survives retrieval
As mentioned above, it starts at a very affordable price point.	PromptScout plans start at $15 per month. Provider-backed monitoring requires choosing a plan before the first run.
Our integration options are extensive and flexible.	PromptScout connects to Google Search Console, Bing Webmaster Tools, and Cloudflare, and exposes an MCP server for agent access.
Many teams find it works well for their reporting needs.	Agencies use the weekly report to show one client which AI answers changed, which competitor gained ground, and which sources were cited.
It's great for growing companies of all sizes.	PromptScout supports recurring buyer-question monitoring across five providers. Check current plans for prompt limits, cadence and access to Opportunities.

The rewritten column gives readers facts they can check. The KDD 2024 GEO benchmark tested content changes, including statistics, quotations and citations, and reported improvements under its evaluation setup. That is research evidence for testing useful specifics, not a guaranteed effect on current assistants or this site. Never invent a number to make a sentence quotable.

Seven habits that make the difference:

• Lead with the answer in one sentence, then justify it. Not the other way around.
• Name the entity in the sentence. Do not let a pronoun do work that points back at the H1 three screens up.
• Include the number, date, currency and unit when you have them. Keep the source and method for outcome claims. An accurate qualitative fact is better than an invented metric.
• Attribute the facts you borrow, with a link. Let readers inspect the evidence and its limits.
• Answer the awkward questions too. Explain fit, limits, setup effort and cost when the evidence supports them. These details help a reader choose; they do not guarantee a recommendation.
• Group each section around a reader question. Use subheadings when they make a complete explanation easier to follow.
• Write a useful summary near the start. Keep essential qualifications beside the answer. There is no magic first-100-words rule.

Choose a page type for the buyer decision

The right format depends on the question and the evidence you can contribute. The options below are not ranked by expected citations or return on effort.

Honest comparison pages

A comparison can help when the buyer needs to choose between specific alternatives. Publish current criteria, evidence, tradeoffs and ownership disclosure. A vendor-authored comparison is not independent endorsement, and no format has a universal advantage over a useful explanatory guide.

Show how the decision changes with the buyer's constraints. For example, an illustrative comparison might say:

Tool A fits a small engineering team that values a short setup. Tool B fits a team that needs reporting across several departments. Verify each option against the same criteria before deciding.

For a real comparison, replace the illustrative tools with dated, verifiable product facts and explain when your own product is not the right fit.

Pricing pages with actual prices

If you can publish prices, give the currency, billing period and material limits. If pricing is bespoke, explain the inputs to a quote without inventing a typical range. Make the buying process clear even when an exact price is unavailable.

Integration and compatibility pages

Compatibility questions need concrete answers: what connects, what does not, and what setup requires. Create a separate integration page only when it adds distinct useful information; otherwise improve the relevant existing documentation.

Documentation

Public documentation can answer setup and compatibility questions precisely. Keep important facts accessible where appropriate, while preserving required authentication and privacy controls. A private document is not a public search source.

Use-case and job pages

Organize these pages around a real problem and supported constraints. A generic answer with an industry name swapped in adds little; a distinct decision guide may be worth its own route.

What performs worse than teams expect

• Essays without useful evidence. Add an original explanation, method or attributable observation.
• Press releases without reader value. Explain the facts and implications instead of relying on promotional claims.
• PDF-only evidence. Check access and provide an accurate HTML summary; formats and crawler behavior vary.
• Gated evidence. Publish a useful public explanation when appropriate; do not remove required access controls merely to pursue citations.

Where else you need to exist

Your own site is necessary and not sufficient. Engines describe you using whatever they can find, and much of it was written by other people.

Our May 28–August 25 panel recorded Reddit at 27.5% of returned ChatGPT source links and between 0.2% and 4.3% for the other engines. These are historical source patterns, not estimates of what a new community post would achieve. Inspect recent sources for your own questions before choosing legitimate, relevant community participation or editorial work.

Before commissioning any outreach, know that there is no fixed list of sites to get on. This is the distribution of returned source links across the domains our panel recorded.

What the long tail means for you. Roughly half of returned source links came from outside the hundred most frequent domains, and more than half of returned domains appeared exactly once. This describes our panel; it does not predict which placement or content format will earn a citation for your brand.

The off-site work that holds up

In rough order of durability:

1. Correct the basics wherever they are already published. Your category, what you do, who it is for, what it costs. Start with your G2 and Capterra profiles, your Crunchbase entry, your LinkedIn company description, and any "best X tools" roundup that already lists you. A stale description is worse than no description, because it is confidently wrong.
2. Earn coverage that describes you in buyer language. A write-up that uses the words your customers use is more retrievable than one that uses the words your brand guidelines prefer.
3. Take review platforms seriously. They are structured, comparative, and constantly updated, which is a retrieval-friendly combination.
4. Answer questions in public, as yourself, where your buyers actually ask. Community answers are read as evidence by at least one major engine. Do this transparently; astroturfing is both dishonest and, given how these systems weight consistency across sources, unreliable.
5. Keep your entity clear. Consistent name, consistent description, consistent category across everywhere you appear. Ambiguity about what you are is a bigger obstacle than obscurity.

Things that are widely sold and do not work

AEO attracted a lot of confident advice very quickly. Some of it has since been tested. Here is what held up and what did not, so you can stop paying for the second column.

The claim	What the evidence says
Publish an llms.txt file as a citation shortcut	Google does not require new AI text files. A documentation map is not proof of a citation benefit.
Add AI-specific schema markup	Google states there is no special schema.org structured data needed for AI Overviews or AI Mode, and no new machine-readable files or markup.
There is a separate AI ranking system to optimize for	Google says its AI features are grounded in the same core ranking and quality systems as Search, with no additional technical requirements.
Chunk your pages into AI-friendly fragments with special delimiters	Retrieval systems do their own chunking. Writing self-contained sections helps. Inventing a fragment format for machines that never asked for one does not.
Repeat your brand name to force an association	Repetition alone is not evidence of useful information or improved recommendations.
Buy placements on "the sites AI cites"	The top hundred domains covered under half of returned source links in our panel, and the list is unstable quarter to quarter.

One clarification on structured data, because that row gets over-corrected. Schema markup is still worth publishing: for rich results, for making your entities unambiguous, and for keeping machine-readable facts in sync with visible ones. It is not an AI visibility lever on its own. Publish it because it describes your page accurately, not because you expect a citation in return.

A fair test for any new tactic. Before adopting something, ask three questions. Who published the evidence, and do they sell the solution? Was the effect measured against a control, or observed after the fact and narrated backwards? Would it still be worth doing if no AI engine existed? A tactic that fails all three is a hypothesis wearing a best-practice costume.

What to do first, if you only have a month

Ordering matters more than completeness. Cheap, high-certainty work first, which is also the order that gives you something to show before the budget conversation.

Week 1: Make yourself retrievable

Audit robots.txt for AI user agents and fix anything that blocks a search crawler. Confirm your top ten commercial pages render server-side. Check snippet directives are intentional. Remove any consent wall or interstitial standing in front of key content.

Done when: every page you would want cited returns its key sentences in a raw curl.

Week 2: Establish a baseline

Pick 20 to 40 real buyer questions in the language buyers use. Run them across the engines your buyers actually use. Record, for every answer, whether you appeared, who appeared instead, and what was cited. Do not change anything yet.

Done when: you can state your mention rate per engine and name the five questions you most want to win.

Week 3: Repair what the answers already reach for

Look at which of your pages were cited and which competitor pages were cited instead. Rewrite the sections that read badly in isolation. Add the missing specifics: prices, versions, limits, integration names. This is editing, not commissioning.

Done when: every page in your cited set leads with a quotable, specific sentence.

Week 4: Fill the single largest gap

Publish the one page for the sub-question nobody in your category answers well. Usually that is a comparison, a constraint page, or an honest "who this is not for." Start only the off-site work your own citation data justifies.

Done when: the gap is covered and the next run is scheduled.

Resist the urge to change ten things at once. Answers move on their own between runs, which is the subject of Part 4. If you change your site, your comparison page, and your community presence in the same week, you will not be able to tell which one did anything, or whether anything did.

Five failure modes worth naming

These show up repeatedly in teams that put in real effort and get little back.

1. Optimizing for the engine your buyers do not use. All the effort, none of the audience.
2. Publishing volume instead of specificity. Twenty generic posts give an engine twenty ways to find the same non-answer.
3. Ignoring a demonstrated technical blocker. Verify that intended public facts are accessible before claiming an editing task solved discovery.
4. Declaring victory on one run. Covered in Part 4, and it is how internal credibility gets spent.
5. Avoiding relevant comparisons. Leave the reader without the criteria needed to make a choice.

Where PromptScout fits in this part

The website audit reports supported checks and sampled-page findings; it is not proof that every item above was checked across your entire site. Opportunities turn supported monitoring evidence into scoped content work with sources and a later watch window. Use the tool-selection guide when deciding which part of this workflow your team needs.

The website audit documentation covers what it checks and how findings are prioritized.

What is next

Part 4 covers knowing whether any of it worked: which metrics mean something, how to design a question panel that produces a trend rather than a series of snapshots, what run-to-run volatility does to your conclusions, and how to write a report that holds up when someone pushes back.