aisearch.marketing
Research briefing · Aug 2026
Retrieval research · August 2026

The shortlist is written before the search runs.

Answer engines changed materially through July and August 2026. They now name candidate brands before fetching a single page, state how fresh they want a result to be, disagree with each other about which sources count, and retrieve content from sites that block them by rule. This briefing sets out the mechanism, then a playbook for acting on it.

Compiled 25 Aug 2026 ChatGPT · Perplexity · Google AI Mode 4 domains probed from NZ v1 draft · small samples
ours measured with our own instrumentation, this week, on live domains field independent field research, not reproduced by us
If you read one thing

Being named in the model's own opening query is worth roughly 33 times more than being fetched during the search. That splits AEO into two separate games with two separate budgets: getting into the shortlist, and winning the citation once you are in it. Most of what agencies sell as AEO only plays the second game, and the second game is the smaller one.

Part one

The mechanism

Seven things that changed, in the order they affect you.

Finding 01

The shortlist is decided before anything is fetched

When someone asks an answer engine for a recommendation without naming brands, the model writes the first search query itself. That query already contains brand names. In one census, 21 of 27 conversations had brands in the opening query that the user never typed, across 11 of 13 product categories tested.field

The consequence is the single most important number in this briefing.

Named in the model's own query
reaches the answer
68.9%
Merely fetched during the search
2.1%
Roughly a 33x advantage. In the same census there were 86 instances where a recommended brand's website was never fetched at all. The model did not need to read them to name them.field

Read that second line carefully. Doing everything right on your website, getting crawled, getting fetched, and still not being in the model's head, buys you a 2.1% chance of appearing. That is the ceiling on pure technical AEO.

When pre-injection happens, and when it does not

The shortlist is pre-written whenIt is not when
The ask is for a recommendation with no brand namedThe user names specific brands
The model has to generate candidates itselfNo web search is needed at all
The intent is comparative: best, alternatives, instead ofIt is an open complaint with no buying intent

What it means. Your highest-value queries are exactly the ones where the shortlist is pre-written, because "best X in Y" is comparative by definition. For those, the work that moves the needle happened months ago and happened off your website. That is uncomfortable to sell and it is what the data says.

The second filter, after retrieval

Of 3,554 retrieved pages in that census, only the top 110 were cited. Position inside the grouping for your own domain decides which of your pages survives.field

5.2%1st in domain group
4.6%2nd
0.3%6th or lower

If you have six near-duplicate pages targeting one intent, five of them are competing against the sixth for a slot at roughly a fifteenth of the odds. Consolidation is not tidying. It is the highest-leverage on-site fix available.

Finding 02

The search call became a command language

Until mid-August, a search issued inside a ChatGPT answer was a JSON object. Four days later, on the same account, it was a flat pipe-delimited line.field

// 16 August {"system1_search_query":[{"q":"site:intercom.com Fin AI Agent pricing 2026"}]} // 20 August fast|Intercom Fin AI agent pricing 2026 live chat support|30|intercom.com
fast|Intercom Fin AI agent pricing 2026 live chat support|30|intercom.com
Call typeWhich retrieval vertical. Not always the web.
QueryThe expanded query, entity-loaded, often carrying brand names the user never said.
FreshnessMaximum acceptable age of a result, in days.
DomainOptional. Where it expects the answer to live.
Call typeWhat it doesCan you win it with content?
fastWeb search, successor to the old fan-outYes the main event
productCatalogue lookup, names semicolon-separatedFeed merchant data, not pages
businessPlaces lookup with a location parameterProfile maps entity, not pages
imageImage searchPartly rarely audited
genui_runRenders a widget from JSON argumentsNo nothing is citable

There is also a response-length directive taking long, medium or short. Fewer sentences means fewer claims, and every claim is a citation slot.

Two parser breaks in five weeks

On 21 July the pipe label field result_source was replaced by search_result_groups, and the pipe labels were dropped.field On 20 August the search_queries field that mirrors the fan-out disappeared entirely.

Any tool reading the old fields now returns an empty result rather than an error. A blank report reads like a clean bill of health, which is the more dangerous failure. If you use any AEO tool that shows a fan-out, check it is not quietly showing you zero.

Finding 03

Freshness is now an explicit number that varies by intent

The third field is a day window, and observed values track how fast the underlying facts move. Content age stopped being a soft signal you argue about and became a filter with a number on it.field

Stock prices
2 days
Sports results
7 days
Pricing & comparisons
30 days
Earnings guidance
90 days
Reddit, typical
365 days
Reddit, deep
3,650 days
Log-scaled. A decade-wide window on community threads sits beside a two-day window on prices, in the same system.

What it means. A comparison page untouched for 30 days can fall outside the window a pricing query is willing to look at. Not outranked. Not considered. That converts "keep it updated" into a schedule with dates on it, which is the version a client can actually be held to. The inverse matters too: a ten-year window on community content means a 2019 thread mentioning you is still live inventory.

Finding 04

Retrieval and credit have come apart

ChatGPT issues subreddit-level queries by name, with the widest window in the system:

fast|site:reddit.com/r/whatnotapp seller tips Whatnot live selling|3650|reddit.com

What gets retrieved and what gets credited are now very different numbers, and the gap moves by query type.field

Niche community query

Seller tips, live-selling marketplace

Pages fetched71
Reddit threads retrieved48 (68%)
Citations given8
Citations to Reddit6
Expertise is genuinely concentrated in the community, so it is both read and credited.
Vendor category query

Best live chat support software

Pages fetched221
Reddit threads retrieved84 (38%)
Citations given11
Citations to Reddit0
Eighty-four threads read, none credited. Reddit shaped the verdict, vendor pages took the credit.

Reddit's overall citation share fell from 3.83% to under 1% in mid-August while retrieval held. That is a filter applied after the fetch, not a loss of relevance.field

Influence without citation is the measurement problem of this cycle

A source can shape an answer while appearing nowhere in it. Every visibility score built on counting citations, ours included, reads that source as absent. The number is not wrong. It is answering a narrower question than the one the client is asking.

Finding 05

The engines disagree, and the disagreement is exploitable

Perplexity is not a variant of ChatGPT. It picks sources on different rules, and the biggest split is the one nobody optimises for.field

Perplexity

Cites video at scale

YouTube citations, product query38
YouTube citations, how-to22
Reddit threads retrieved3
Reddit citations0
It quotes video content directly, so a video is a first-class citable source.
ChatGPT

The exact inversion

Reddit threads retrieved48–84
Reddit citations, niche6
Video transcript extractionno
Search always runs?no
It cannot extract transcripts, so video is close to invisible to it.

Two more Perplexity mechanics worth knowing. skip_search is always false, so every query hits the web and every query is therefore winnable, which is not true of ChatGPT. And its fan-out is conservative, using the words people actually type rather than a cloud of adjacent topics, so exact-phrase targeting still works there.

What wins, by intent, on Perplexity

IntentWhat takes the citationsYour play
CommercialFresh current-year listicles. Brand size is not the gate, freshness is.Get onto the "best [category] [year]" lists and keep them current
ComparisonThe named vendor's own pageOwn your comparison pages, they are winnable
NewsThe official primary source. Freshness alone is not enough.Be the primary source or do not bother
LocalMaps place-entities take everything, roundups get nothingFix the maps entity, not the content
How-toVideo, cited like any text sourcePublish the video, not just the post

Perplexity also carries scoped trust entries per domain, at level 1 (credible) or level 2 (trusted), and the scope is first-party only. A domain is trusted about its own products and services, not about the category. That is a clear instruction: the page about your own thing is the one you can win, and the "ultimate guide to the industry" is the one you probably cannot.

Finding 06

Three layers of access, three different answers

We probed four domains from New Zealand on 25 August, testing fourteen crawler user agents against robots.txt and then against the live server.ours The results say something the field research implies but does not state outright.

DomainVerdictrobots.txtLive server
reddit.comBLOCKEDDisallow: / for all 14 agentsMixed, see below
gerrards.co.nzOPENAll allowedAll 200
aisearch.marketingOPENExplicit AllowAll 200
janprocleaning.co.nzOPENAll allowedAll 200

Reddit is the case that proves the point

Its robots.txt refuses every AI crawler we test, including OAI-SearchBot. Its live server then disagrees with its own robots file, and disagrees per agent.ours

Agentrobots.txtLive fetchReading
Real browsern/a200Control
OAI-SearchBotDisallow200Server allows what robots forbids
ChatGPT-UserDisallow403Refused twice over
PerplexityBotDisallow200Server allows what robots forbids
BingbotDisallow403Hard refusal
python-requestsn/a200An anonymous scraper walks straight in

Three things fall out of that table, and all three are practical.

The honest limit, and why it cuts both ways

Our probe proves the observable, client-controlled gates only: robots.txt directives and WAF behaviour toward named agents. It cannot see residential-proxy traffic and never claims to.

Which means: a BLOCKED verdict is a real, fixable problem you should fix first. An OPEN verdict is not proof you are being retrieved, and a block is not proof you are not. Anyone selling "we made you visible to AI" off a robots.txt edit alone is overselling a gate they can only half see.

The finding that matters commercially: three of our four domains were already wide open. For a typical NZ business, reachability is not the problem. Selling a robots.txt fix to a client whose door is already open is selling nothing. The problem is upstream in the shortlist, or downstream in the citation.

Finding 07

Google AI Mode conversations leak into Search Console

Fragments of AI Mode conversations are recorded as ordinary queries in the performance report. Not summaries, the literal strings people typed mid-conversation.field

1,127fragments classified
20,300impressions, 16 months
13clicks from 8,834 impr.
57.7%impressions anonymised

They are recognisable because nobody types them into a search box. The tells are conversational, not topical.

SignalLooks likeWhy it cannot be a search
Assistant instructiongive me step by stepAddressed to a listener, not an index
First-person contexti am using lmstudioPointless to tell a search engine
Dangling pronounis it freeThe referent exists only in a prior turn
Politeness markerplease clarifyNobody says please to an input field
Bare replyyes · sure · really?An answer, not a question

Seven categories in total: bare replies, pivot follow-ups such as what about resend?, conversational questions, synthetic tracker probes, pasted agent instruction sets, pasted error strings and CSV headers, and long uncategorised strings.

The two worth money

Access reality, verified There is no AI-specific column in the BigQuery export, the SearchAppearance dimension returns nothing AI-related, and the SearchAnalytics API type parameter stops at googleNews. The UI export is the only route to the Generative AI report. Anyone claiming an automated pull of it is pulling something else.field

The floor problem caps how far this goes. Conversational strings almost never repeat, so most sit at one to three impressions, and 57.7% of impressions are anonymised rows with no string at all. Treat it as a qualitative sample with real volume behind it, not a keyword list.

Part two

The playbook

What to actually do, in the order that pays. Every play below traces to a finding above.

Start here: it is two games, not one

The 33x gap in Finding 01 means these are separate budgets with separate timelines. Running the second without the first is the most common and most expensive mistake in AEO right now.

Game one · the bigger one

Get into the shortlist

Be a brand the model names before it searches. This is reputation across the open web: mentions, reviews, listicles, directories, community threads, analyst and press coverage.

It is slow, it is mostly off your own website, and it is not technical SEO. It is the one that carries the 68.9%.

Wins when your name appears in the query the model writes for itself.
Game two · the faster one

Win the citation

Once you are in the retrieved pool, be the page that gets bound to a claim. Page consolidation, answer placement, server-side rendering, freshness cadence, first-party scope.

It is fast, it is fully under your control, and its ceiling is 2.1% if you are not also playing game one.

Wins when a sentence in the answer points at your domain.

Diagnose before you prescribe

Five failure states. Each has a different test and a different fix, and they are ordered: fix the lowest one that is failing before touching anything above it.

State 0 · rarely the problem

Not reachable

robots.txt or the WAF refuses the crawlers that feed the engines. Real, but uncommon: three of the four domains we probed were already wide open.

TestRun the reachability probe. A BLOCKED verdict outranks every content fix on this list.
State 1 · the expensive one

Not in the shortlist

Your name never appears in the query the model writes for itself. You are competing for the 2.1% lane no matter how good the site is, and nothing in a rank tracker or an ads report will show you this.

TestAsk the engine a clean "best [category] in [place]" with memory off. If the answer names five brands and none are you, and your site was never fetched, this is your state.
State 2

Named, but never fetched

The model has you in mind but does not go to your site, or goes to the wrong domain. Entity confusion, a stale domain slot after a rebrand, or a name that collides with a bigger entity.

TestYou appear in the answer but your domain is absent from the retrieved pool. Check the domain slot in the search call, and check for name collisions.
State 3 · the cheapest fix

Fetched, but buried in your own domain group

Your pages are being retrieved and then filtered out. Sixth position or lower in your own domain grouping is a 0.3% citation rate against 5.2% for first.

TestCount how many of your pages target the one intent. If it is more than one, they are competing with each other, and consolidation is the single highest-leverage on-site move available.
State 4

Surfaced, but not cited

You are shown as a source and no claim is bound to you. Usually the answer to the question is not present as a liftable sentence, or it sits too deep in the HTML, or the page is client-rendered.

TestRead your page as the model does. Is there one clean sentence that answers the question, high in the source, present without JavaScript?

Game one plays: get into the shortlist

  1. Find the subreddit that carries your category, then read it. One community usually dominates, and a single subreddit will often supply a third or more of the threads on a category query. Count who gets named and how often. That list is the shortlist the model is learning from, and it is frequently not your paid-search competitor set.
  2. Get onto the current-year listicles. On commercial intent, fresh "best [category] [year]" roundups take the citations and brand size is not the gate, freshness is. This is outreach work with a deadline attached, and it has to be redone annually.
  3. Treat aggregator and directory lists as a primary route, not a support tactic. Our own client data shows a business ranked second in an answer through a third party's list, with zero citations to its own domain, while a larger competitor with a stronger site was fetched nine times and neither named nor cited. Riding someone else's list is a real, valuable, revocable position. Take it.
  4. Be consistently named. The model matches brand tokens. If you appear as three variants across the web, you are diluting the one signal that gets you pre-injected. Pick the string and enforce it everywhere.
  5. Accept the timeline. This is years of open-web presence, not a quarter of content. Say so up front rather than selling a 90-day version of it that cannot exist.
The honest caveat to lead with Community and PR work is upstream of the citation and is not directly attributable. You will not see it in a rank tracker, and the engine may read a thread that mentions you and credit a vendor page instead. Measure it as share of mentions in the corpus, not as citations.

Game two plays: win the citation

  1. Consolidate to one page per intent. The 5.2% versus 0.3% spread between first and sixth in your own domain grouping means near-duplicates are actively costing you. Merge, redirect, and leave the single best page standing.
  2. Put the answer sentence early and make it liftable. The citation binds to a claim. A clean declarative sentence that answers the question, high in the HTML, is what gets bound. Burying it under 800 words of preamble is a choice with a cost.
  3. Render server-side. Common Crawl's crawler does not execute JavaScript, and it is not alone. If your content only exists after hydration, a meaningful share of the pipeline sees an empty page.
  4. Set a freshness cadence by content type, with dates. Use the observed windows: pricing and comparison pages inside 30 days, anything financial inside 90, evergreen guides on a slower loop. This is a schedule, not a sentiment.
  5. Own the comparison page for your own product. Trust is scoped first-party. You are trusted about your own thing and not about the category, so "our product versus theirs" is winnable and "the ultimate guide to the industry" mostly is not.
  6. Publish the video where video wins. Perplexity cites YouTube at scale on product and how-to intent, 38 and 22 citations respectively in one sample, while ChatGPT cannot extract transcripts at all. If your buyers use Perplexity, a video is a citable asset. If they use ChatGPT, it is close to invisible. Pick per audience.
  7. Check your CDN is not blocking on your behalf. A default CDN rule added roughly 3.8 million domains to the blocked pile. Many owners have no idea. This is a five-minute check that occasionally finds a catastrophe.

Per-engine cheat sheet

The engines reward different assets. Optimising for one average of them all is how you end up strong nowhere.

 ChatGPTPerplexityGoogle AI Mode
Does every query search?NoYes, so every query is winnableVaries
Community threadsRetrieved heavily, credited on niche intent onlyRetrieved, cited close to neverPresent in the index
VideoCannot extract transcripts, weakCited at scale on product and how-toSurfaces separately
FreshnessExplicit day window in the callCurrent-year listicles win commercialNot exposed
Local intentSeparate places lookupMaps entities take everythingMaps-led
What you can measureFull retrieval funnel from the conversationStreamed router scorecardLeaked fragments in Search Console
Your first moveGet named in the pre-search shortlistGet on the fresh listicle, publish the videoMine pivot follow-ups for content gaps

The audit you can run this week

Ten checks. Everything here is either a command we already have or a report you already have access to.

1. Probe reachability, per agent

Not a single yes or no. Allowing the search crawler while refusing the live fetch is a specific and currently invisible failure.

node scraper-access-check.mjs yourdomain.co.nz

2. Check the CDN is not blocking for you

Default bot rules block AI crawlers without the owner ever choosing it. Look at the rule, not the intention.

3. Ask the engine cleanly and watch the shortlist

Memory and custom instructions off, or a clean profile. Ask "best [category] in [place]". Write down every brand named. That is the shortlist you are trying to join.

4. Pull the community corpus and count names

Find the dominant subreddit, retrieve the threads, count brand mentions. Your share of that count is your game-one scoreboard.

node lib/run.js reddit-sentiment --query "your category" --limit 25 --comments 5 --hide

5. Count your near-duplicate pages per intent

More than one page for one intent means they are competing with each other at a fifteenth of the odds. This is usually the fastest win on the list.

6. Date-stamp your commercial pages

Anything with a price or a comparison on it, last modified more than 30 days ago, is outside the window for pricing queries. List them, schedule them.

7. View source with JavaScript disabled

If the answer sentence is not in the raw HTML, a meaningful share of the retrieval pipeline never sees it.

8. Filter Search Console for conversation fragments

Search results, add filter, query, custom, regex. Then read the pivot follow-ups as a content brief and bin the machine probes.

^(yes|yeah|ok|okay|sure)[?!.,]*$

9. Verify your AEO tool is not silently empty

Two field renames in five weeks. A tool reading the old fields returns zero, and zero looks like a clean result. Confirm it is reading the current format.

10. Record which engine your buyers actually use

The plays diverge sharply between ChatGPT and Perplexity. Optimising for the average of both is how you end up strong on neither.

Part three

How much weight this holds

Stated plainly, because the alternative is selling a number we cannot defend.

The through-line

Retrieval and credit have come apart, and the decision that matters now happens before retrieval starts. Systems read far more than they cite, they are explicit about how fresh a page must be, and they arrive at a shortlist built from years of open-web reputation rather than from your website. Measurement built on citation counting reads a shrinking slice of that, and will keep returning confident numbers about a narrowing question.