The AI Search Operating Manual

What Atlas measurement has found to date

Atlas ran 8,209 prompts across 11 AI engines for 16 brands from July to September 2026. These six findings recur through the manual.

• 17.3%of prompt and engine pairs changed whether the brand was mentioned across repeated runs of the same prompt on the same engine.Part 02
• 2 to 5searches per answer, the median for AI engines that show the searches they ran.Part 02
• 89.3%of citations pointed to websites the brand does not own.Part 06
• 5.5xBrands were mentioned 5.5 times as often as they were recommended on category prompts.Part 12
• 20xOwned articles updated in the last 90 days were cited about 20 times as often per article as articles older than a year.Part 04
• 3xFirewalls and CDNs refused AI crawlers about three times as often as robots.txt blocked them.Part 09

Atlas measurement, 16 brands, 11 engines, 2026-07-13 to 2026-09-14. Each section states its finding's exact window. Brands are not named.

How much to trust each recommendation

The recommendations in this manual rest on four tiers of evidence. The tier decides what a recommendation is allowed to claim. Tactics with no evidence at any tier are listed in the myth section (part 15) instead of recommended.

Brands missing from training data start 57 points behind

Published tests of invented brands against real ones report roughly a 57-point gap on a 100-point reranker scale. A brand absent from training data starts that far behind, and no on-page work closes that gap; only presence in the public record does.

Published tests of invented brands against real brands on a 100-point reranker scale.

Search and AI answers share the same first six steps

The same six stages run before either product renders anything, and a page can be excluded at any one of them.

1. 01Query understandingThe system converts the prompt to an embedding before it retrieves anything.
2. 02Query transformationThe system rewrites the prompt and expands it into many search variants (query fan out).
3. 03Multi-path retrievalThe system retrieves candidates along more than one path at once.
4. 04Candidate processingThe system converts retrieved items into candidates it can compare.
5. 05Multi-stage rankingCheap rankers run first, and expensive rankers run only on the survivors.
6. 06Eligibility and filteringFilters remove candidates the product is not allowed to use.

The search path and the grounding path diverge after stage 06.

After that, search and AI answers split

After filtering, the same candidate pool feeds both products, and each product applies its own selection rules.

A page tuned for the results page can be excluded from evidence selection, and the reverse.

AI engines run 2 to 5 searches per answer

When an engine shows the searches it ran, Atlas records them. Perplexity ran a median of 5 searches per answer; Gemini with web search ran 3; ChatGPT and Claude with web search ran 2. A page written for one phrasing matches only one of those searches.

An earlier draft of this manual cited about 15 search variants and 600 candidates per prompt. Atlas could not verify that figure, so the chart shows only the searches engines exposed.

Atlas measurement, 16 brands, 4 engines that expose their searches, 49,462 answers, 2026-07-27 to 2026-09-14. The other 7 engines do not expose searches.

You cannot fully observe why and how AI systems rank your pages

Publishers cannot measure the ranking stages directly, so find the gate a page does not pass before applying any checklist.

1. 01Retrieval loopsEvidence selection can reject the whole retrieved pool and send the system back for more.
2. 02Ranking cascadesCheap rankers run first, and expensive rankers run only on the candidates that survive them.
3. 03PersonalizationSession memory and personalization give two users different candidates for the same prompt.
4. 04No outside viewPublishers cannot observe candidate sets, reranking scores, or context allocation. Answers to one prompt change by model, product mode, location, date, and random variation, so one screenshot proves nothing.

Atlas measurement, 16 brands, 11 engines, 29,563 prompt and engine pairs with 2 or more runs, 2026-07-13 to 2026-09-14.

A page must pass four checks to be cited

Gates are pass or fail. A page that does not pass one gate is excluded from the answer, regardless of the other three.

1. 01DiscoverableCan the system reach it?Crawlable URL, sitemap, no robots block, clean canonical. Also check for a firewall or CDN that challenges the bot by IP while robots.txt allows it.
2. 02ParsableCan the system extract structure?Semantic HTML, valid structured data, alt text, transcripts, clean heading hierarchy. A visually impressive JavaScript blob with no semantic structure does not pass this gate.
3. 03GroundableCan a passage be trusted on its own?Self-contained claims, named sources, no contradictions on the page. Grounding stages deprioritize ambiguous passages.
4. 04TrustworthyDo other sources and time agree?Entity data that matches public knowledge graphs, attributed quotes, and the same claim stated on several credible websites over time.

A gate is pass or fail, never a score.

Recently updated articles were cited far more often

Atlas matched owned articles to the citations they received. Articles updated in the last 90 days were cited about 20 times as often per article as articles older than a year. For a recency-sensitive task, stale information is ineligible regardless of quality, and the system decides what counts as recency-sensitive. Track fact freshness, not the date stamp.

Two signals most teams skip: submit updated URLs through IndexNow, and watch access logs for 499 status codes, the signature of an AI crawler abandoning a slow response.

Atlas measurement, 13 brands, 5,092 owned articles with a published modified date, 2026-07-13 to 2026-09-14. Observational: frequently updated pages may also be the pages brands consider most important.

Cite your sources so AI systems can verify your claims

Verification stages check named sources, specific data, and traceable claims against each other. A stage can drop a whole passage when one of its claims is contradicted.

"Studies show" is unverifiable. A named source with a year is evidence.

AI answers quote single passages from indexed pages

Retrieval selects whole pages. Answer generation then quotes one passage from a selected page. The page is the unit of indexing; the passage is the unit of citation.

Models weight the start and end of a passage most, so an introduction placed before the answer lowers its chance of selection. Put the answer first and the evidence after it. Semantic richness makes content matchable and extractable: three thousand words under eight precise headings can be cited eight times across eight queries, while the same words as a wall of prose extract nothing.

A direct answer is good for people and for AI.

A brand is the average of everything it publishes, and everything public said about it

In vector space that average is called a centroid. Models build their picture of your brand from that text: the system calculates your brand rather than seeing it.

Each page is one point. A brand cited across many prompts covers a connected region of related topics.

How brands become interchangeable

A centroid far from competitors' centroids makes the brand easy to tell apart. Two failure modes erode that: cluster collision and centroid drift.

Each small dot is one piece of published content. The large dot is the average. These drawings are schematic and are not plotted from measured data.

How AI selection criteria have changed

The criterion that decides selection changed twice: presence first, then similarity to the query, now connectedness of evidence.

Inputs that shape an AI answer

Models combine several sources into one answer. Brands can influence most of those sources, and Momentic calls that set the Influenceable Web: the brand's website, its owned content, its reputation surfaces, offsite and social coverage, the training data already inside the models, and the prompt itself.

The sixth input is the prompt itself. Its sentiment influences answer sentiment, and the entities it treats as central decide which brands the model includes.

Places of corroboration

On these eight kinds of websites, a third party can confirm a brand's claim. Each confirmation starts with evidence the brand publishes.

Atlas measurement, 16 brands, 1,288,263 captured citations, 2026-07-13 to 2026-09-13. Owned share by engine: 8 engines that cite sources.

Communities are where first-hand experience genuinely exists. Respect that in your strategy.

Tactics Momentic does not use

Momentic excludes four tactics outright and builds press on real datasets, qualified experts, and first-hand results instead. The tactic class carries platform policy risk and reputational risk.

1. 01Purchased mentionsPayment for coverage that is presented to readers as independent.
2. 02Staff sockpuppetsEmployees posting as unaffiliated customers or unaffiliated experts.
3. 03Seeded forum threadsDiscussion planted by the brand and staged to read as organic.
4. 04Listicle insertionA ranking position bought inside a third-party roundup.

Sources for demand research

A keyword is basically a topic. A prompt adds constraints, context, and the outcome the person wants. Demand research now runs three layers.

1. 01Search demandWhat people already type into a search box, counted in keyword volume and internal site search.
2. 02Decision questionsWhat a buyer needs answered before acting, taken from sales calls and support tickets.
3. 03Query fan out coverageThe sub-questions a system retrieves separately when it answers one visible prompt.

Validate the panel against keyword data, internal site search, sales calls, and support tickets. Group recurring sub-questions into entities, comparisons, criteria, risks, procedures, and next actions; map each cluster to the strongest existing URL; create a new URL only for a genuinely distinct task. Do not build a page per paraphrase: platform guidance is explicit that scaled low-value variants are the wrong answer.

Prompt classifications

A panel is a fixed, versioned, representative set of prompts per commercial journey, preserved across months.

Before comparing, do this

Wording and denominators decide whether a panel's numbers can be compared across brands or months. Run both checks before any comparison.

Each rate names its own denominator

Report each rate next to the denominator it was divided by, in the same sentence.

Six kinds of material competitors cannot copy

If a model can write your article from ten similar pages, it has no reason to cite yours. Publish material competitors cannot copy without doing the same work.

Write sections AI can quote

A model quotes one section under one heading as a unit, so each section must resolve its own question completely. Models do not select passages by word count, so do not split sections to meet a word target.

1. 01Answer immediatelyThe direct answer appears in the first sentences under the heading, before context, caveats, or background.
2. 02Qualify scope and dateLabel the date, the units, the method, and where the numbers came from, inside the same section.
3. 03Evidence the claim"Studies show" cannot be checked. Name the source and the year so a model can verify the claim.
4. 04Explain why it mattersState the consequence next to the claim, so the passage still makes sense read away from the page.
5. 05Connect to the next questionThe next sub-question gets its own heading, never a new page written to repeat the same answer.

Write each section so it makes sense when a model quotes it alone.

Update facts on the cadence the topic actually changes, distinguish published, reviewed, and data-through dates, and retire dead URLs cleanly. Where the task is visual (product, local, how-to, comparison), imagery and video are answer inputs: transcripts, captions, descriptive filenames, and surrounding text make them legible, and a fact that exists only inside an image does not exist.

What the 40% study measured

The 2024 Princeton and KDD GEO paper supplied top search results to the generator first, then changed the content. The 40% figure applies only inside that setup.

1. 01SupplyThe paper gave the generator the top search results before it wrote an answer.
2. 02RewriteOne of those supplied sources then had its content changed.
3. 03ScoreThe changed source was scored on how visible its attributed text was in the answer.

Do not repeat the claim that adding statistics increases ChatGPT rankings by 40%. It collapses retrieval, context allocation, citation, and traffic into one imaginary metric. The honest reading of the research: quotations, statistics, and sourced claims can change visibility within a fixed context, while keyword stuffing does not, because concrete, attributable information is easier to reuse than vague prose.

The 2024 Princeton and KDD GEO paper; the 2026 critical survey of 45 GEO studies.

Which AI crawlers to allow, and why

Search inclusion, model training, and user-invoked agent access are three separate decisions, and each system below publishes its own user-agent tokens. Set the policy per decision and per system, and do not copy a generic blocklist.

Check crawler access at four layers

Robots.txt records only the stated policy. A firewall or a CDN rule can refuse a crawler that robots.txt allows, and only the access log's error rate shows that the refusal happened. The audit reads the four layers in order until one refusal is found.

1. 01Robots.txtWritten per purpose, because search inclusion, model training, and agent access are separate decisions.
2. 02Firewall rulesRobots.txt allows the bot while the WAF challenges it by IP, and nothing in robots.txt reports the refusal.
3. 03CDN bot managementManaged bot rules can block what robots.txt permits, so WAF and CDN policy have to match robots intent rather than contradict it.
4. 04Access-log error ratesThe only direct measurement: which crawlers came, what they fetched, what they got back, and whether they gave up.

Atlas bot-reachability monitor, 16 brands, 98 domains, latest snapshot, 2026-09-14. Refused = HTTP 403 on a live probe.

Technical checks

A crawler must fetch the page, a renderer must see its content, and the index must keep it. If a page fails one of these seven checks, AI-specific work will not fix it.

1. 01Stable 200s and clean canonicals
2. 02Important content present in rendered text
3. 03Internal links from related pages, plus a hub page per topic
4. 04WAF and CDN policies matching robots intent
5. 05One set of canonical facts across templates and feeds
6. 06Sitemaps correct
7. 07IndexNow where the ecosystem supports it

Server logs are the only direct measurement of the first gate: build bot cohorts, trend them, and alert on error-rate shifts. Do not build a shadow site: a separate "AI version" of the website creates drift and duplicate truth. The primary experience, made semantically clean, is the AI version.

Google: AI Overviews and AI Mode

Google documents how these features retrieve pages, so each check targets a documented behavior.

• Ordinary crawl and index health on priority pages
• Subtopics covered on existing pages, without one page per query
• Original data or research that other pages do not already contain
• Structured data valid and matching visible content, for supported features
• Search Console's Generative AI performance report reviewed monthly (global since August 2026)
• Google-Extended decision documented separately from Search

ChatGPT Search

ChatGPT rewrites a prompt into one or more targeted searches, runs them, and attaches citations to the sources it quotes. Five checks decide whether a page is available to be one of those sources.

• OAI-SearchBot allowed in robots and through the WAF, verified in logs
• Pages indexed and ranking in regular search
• Titles, snippets, and page identity precise enough to match the rewritten sub-query
• GPTBot policy decided separately from search access
• chatgpt.com referrals tracked in analytics, next to panel visibility for answers that send no click

Perplexity, Claude, Copilot and Bing

These engines share the same crawl and index foundations. Each enforces its own crawler access and publishes its own first-party report, so the audit checks the three groups separately.

Six product fields AI shopping agents read

Agents read product data to complete purchases, so identifiers, variants, price, availability, shipping, and returns must match across page, feed, and backend.

1. 01Identifiers
2. 02Variants
3. 03Price
4. 04Availability
5. 05Shipping
6. 06Returns

Rules

• Product copy organized around the criteria a buyer decides on.
• Feed errors fixed with the same priority as website errors.

Four facts to get right for local results

AI engines check a location's name and address, location page, hours, and reviews before they show the business in an answer. Get these four facts right first, on the website and on the brand's listings.

24 checks so AI agents can use your website

Browser agents read the DOM the way assistive technology does. Momentic's agentic-UX audit scores each of these 24 items 0 for missing, 1 for partial or inconsistent, and 2 for deliberate, verified, and maintained.

Momentic does not recommend an agent-operability change that makes the page harder for a human to use.

The Momentic agentic-UX audit instrument. Item IDs and wording as published.

Measure each stage separately

A composite score does not show which gate failed. Each layer reports one metric, and one named system supplies it.

Atlas measurement, 16 brands (each with at least 30 graded category-prompt answers), 2026-07-13 to 2026-09-13.

Four metrics teams often confuse

Each rate changes independently of the other three, so every report states the denominator next to the number.

1. 01Mention rateThe answer names the brand. Denominator: category prompts only; on a branded prompt a mention is a given.
2. 02Citation rateThe answer uses an owned page as evidence. Denominator: the prompts run, with branded and category kept apart.
3. 03Recommendation rateThe answer actively suggests the brand. Denominator: category prompts, where the answer is free to choose.
4. 04Referral rateA visit arrives from an AI surface. Denominator: sessions in analytics, never answers.

Category prompts: Atlas measurement, 16 brands, 11 engines, 88,574 answers to category prompts, 2026-07-13 to 2026-09-14. Owned-citation share: 16 brands, 344,073 citations on branded prompts, 2026-07-13 to 2026-09-13. Caveat rate: 16 brands, 64,282 graded branded-prompt answers, 2026-07-13 to 2026-09-13.

Six rules for reporting results

Check each number against these six rules before publishing it.

1. 01Branded prompts do not measure mentionsMeasure owned-citation share and caveat rate on branded prompts. Mention rate belongs to category prompts.
2. 02Citations do not attribute to revenueBranded search volume and direct visits record influence. The CRM records revenue on its own terms.
3. 03Every number names its denominator where it appearsCompute answer share over the answers to the period's category prompts, never over a brand's own rows.
4. 04Framing balances to neutral firstBalance the wording of a prompt set before running any cross-brand or cross-month comparison.
5. 05Two numbers compare only when their run mix matchesMatch engines, product modes, and panel version, or publish a comparability note instead of a trend.
6. 06One run does not establish an effectReport treated and comparison sets, repeated runs, and the range of the effect.

A number without a stated denominator is not a result.

What to do in the first 90 days

Fix crawler access first, then publish original evidence, then run tests. Finish each window's job before starting the next.

What to start now, test, and skip

Momentic groups the work into four buckets, from what to start now to what to skip. Each bucket lists five items.

Set priorities before setting dates.

The ten parts of the audit

Momentic scores each instrument 0, 1, or 2, records the evidence, and weights each category by business type.

1. 01Technical eligibility and rendering
2. 02AI crawler access across the four layers
3. 03Indexation and architecture
4. 04Demand and panel coverage
5. 05Content value and extractability
6. 06Entity and corroboration
7. 07Structured and feed data
8. 08Platform readiness per engine
9. 09Agentic UX across 24 action-readiness items
10. 10Measurement and governance

Scoring: 0 missing, 1 partial or inconsistent, 2 deliberate, verified, and maintained. Totals are not compared across businesses.

97 of 100 llms.txt files got zero requests

Ahrefs studied 137,000 sites and logged requests to each valid llms.txt file it found. 97% received no request in the measured period.

Google states that AI text files are unnecessary for its generative features. An llms.txt file can help documentation websites, but no published evidence shows it improves rankings.

Ahrefs llms.txt analysis across 137,000 sites; Google Search Central on AI text files.

Twelve claims to stop repeating

Each claim below circulates as a tactic. The correction states what published research or Atlas measurement supports.

1. 01"Add llms.txt to rank in AI."Google states AI text files are unnecessary for its generative features, and Ahrefs found that 97% of valid files received no requests.
2. 02"Schema makes models cite you."Schema is infrastructure for supported features and entity clarity. A tracked study across nearly two thousand pages found citations barely changed after JSON-LD was added.
3. 03"Every answer needs a 40-word chunk."Google confirms there is no required length. Write each section to answer one question completely.
4. 04"The 40% study proves a ranking boost."The study put the sources into the model's context first, so its 40% gain does not apply to live retrieval.
5. 05"Rank #1 and you will be cited."AI answers regularly cite pages beyond the top search results, and the overlap changes by engine and month.
6. 06"Cited means recommended."On category prompts in Atlas data, brands were cited in 14.3% of answers and recommended in 2.2%. Track each separately.
7. 07"Fresh always wins."Freshness matters for facts that change, such as prices. Evidence decides stable facts. Changing only the date does nothing.
8. 08"Reddit is the ranking factor."Real community posts help on some prompts and engines. Fake posts break platform rules and get removed.
9. 09"Traffic no longer matters."Referral counts miss answers that send no click, but the visits that arrive are real. Track answer metrics alongside traffic and revenue.
10. 10"AI search needs its own content team."The work spans search, content, PR, product data, and analytics. A separate team tends to publish duplicate pages with conflicting facts.
11. 11"Write for machines now."Writing that is clear to people is also what models extract best. Google asks for content written for people first.
12. 12"There is one algorithm to reverse-engineer."Each product uses different models, retrieval systems, and modes. Measure each one repeatedly and report ranges.

The manual in ten steps

Run the whole manual in this order.

1. 01Choose the taskPick a task a buyer needs done.
2. 02Check reachabilityThe four access layers, verified in logs.
3. 03Map the query fan outSub-questions, entities, criteria, evidence.
4. 04Own a unique factPublish what no model can synthesize from competitor pages.
5. 05Make it explicitAnswer first, then label dates, units, methods, provenance.
6. 06Earn corroborationGive journalists, partners, and reviewers evidence they can cite.
7. 07Run the panelSame prompts, each engine, framed neutral: mention, citation, recommendation, accuracy.
8. 08Read first-party signalsSearch Console generative, Bing AI Performance, server logs, referrals.
9. 09Test with controlsComparison sets, repeated runs, effect ranges.
10. 10Optimize the actionMake the next step, such as booking or buying, easy for a person or an agent.