Who Sets the Agenda When the Answer Has No Front Page? The Invisible Gates of Answer Engines

Two answers, two worlds

Two people ask an AI assistant about the same protest. The first types: Who is protesting, and why? The second asks: Why has this demonstration brought the city to a standstill? Both receive fluent, matter-of-fact answers. In one, the protesters’ demands occupy the centre of the story. In the other, the leading characters are traffic, losses to local businesses and the police response. Every sentence in both versions could be true. Neither need be a conventional lie. Yet each arranges a different stage: it assigns the actors their roles, defines the problem and decides who is entitled to explain first.

The scene is hypothetical; the mechanism is not futuristic. Answer engines and AI assistants increasingly decline to show us a shelf of material. They bring a plated meal. As a convenience, this is excellent service. As communication, it raises a question that no grammar test can resolve: who selected the ingredients, and how can that selection be challenged?

In a newspaper, the agenda was visible in the form. The largest headline occupied the top of page one, a secondary story sat below it, and other matters did not make the edition. Readers could criticise the newsroom, compare two papers or write to the editor. In a generated answer, decisions are distributed across data access, retrieval, ranking and synthesis. Their

seams are polished away. The user receives a paragraph spoken in a voice without an institutional face.

The practical problem is invisible gatekeeping without an adequate audit trail. A source may shape the answer but never be displayed. A decisive document may fall outside the accessible corpus. A passage may be retrieved while its caveat disappears in compression. Finally, the model’s inference may sound like an established fact. The issue is not only that a publisher has lost a click. It is that the reader cannot see the doors through which this picture of the world has passed.

The front page as a machine for making importance

Agenda-setting is often summarised as the media telling people what to think about. The shorthand is useful, but imprecise. In their classic Chapel Hill study of a US presidential campaign, Maxwell McCombs and Donald Shaw compared the issues prominent in news coverage with those undecided voters considered most important. They did not establish a simple causal law in one study, and carefully noted that limit. They did find a striking correspondence between the media’s hierarchy of issues and the public’s hierarchy of perceived importance (McCombs & Shaw, 1972, pp. 176–187).

Audiences learn ‘how much importance to attach to that issue’, as the authors put it (McCombs & Shaw, 1972, p. 176). Importance need not be declared. Position, frequency, airtime, length and repetition perform the work.

Gatekeeping describes an earlier point in the same process: information passing through gates of selection. A reporter chooses sources, an editor chooses stories, a desk chooses the headline and position, and a platform chooses the order of distribution. The gatekeeper need not be a lonely guard with a key. It is better understood as a chain of decisions, norms and organisational constraints (Shoemaker & Vos, 2009).

An answer engine combines both functions. It first admits some material and withholds the rest; then it establishes importance inside the finished response. The difference lies in the interface. A traditional homepage at least displays its hierarchy as a hierarchy. A seamless answer can imply that there is no arrangement at all, only neutral content. It is as if the menu has vanished

because the waiter immediately brings today’s special. Somebody still decided what to cook. The diner simply never saw the alternatives.

Three gates, not one mysterious algorithm

The word algorithm often functions as stage smoke. Saying that ‘the algorithm selected it’ sounds technical but conveys little more than ‘the kitchen made dinner’. To locate responsibility, we need to distinguish at least three gates.

The corpus gate determines the world of material from which an answer can possibly draw. Depending on the product, that world may include patterns acquired during training, a search index, licensed publications, documents uploaded by the user, or a mixture of all four. A politics of visibility is already operating here. Languages, regions, types of publisher and historical periods are not represented equally. A document outside the system’s reach has not thereby ceased to exist.

The retrieval gate operates when a particular question arrives. The system interprets and may reformulate the query, retrieves passages and ranks them. Two similar prompts need not bring back the same evidence. A document may be inside the corpus yet lose the retrieval contest to an item with better metadata, greater recency or more links. Technical relevance is not identical to public importance. The easiest material to retrieve is not necessarily the best witness.

The synthesis gate assembles the retrieved pieces into speech. Here the system decides which proposition opens the answer, what is compressed, whether a connective word implies causation and whether disagreement among sources remains visible or is ironed flat. Quotation, paraphrase, inference and stylistic glue may all look the same. Algorithmic news production does not abolish editing; it embeds editorial decisions in data, models and system rules (Diakopoulos, 2019).

These gates are not a full technical diagram of every product. Some assistants do not search the live web, and suppliers differ markedly in architecture and disclosure. They are a communication model: ask what could enter, what was retrieved this time and how it became this answer.

Transparency then stops meaning the magical exposure of an entire model. It becomes a request for particular traces attached to a particular claim.

Even a correct answer can become a cul-de-sac

In 2025, the BBC and European Broadcasting Union coordinated an international assessment of more than 3,000 answers from four AI assistants to news questions. Twenty-two public-service media organisations took part across 18 countries and 14 languages. Evaluators found a significant issue capable of materially misleading an audience in 45 per cent of the answers assessed. The catalogue included factual errors, missing context, confused distinctions between opinion and fact, and defects in sourcing (BBC & European Broadcasting Union, 2025; Yezza et al., 2025, pp. 3–8).

This does not prove that almost every second answer from any AI is false. The study concerned specified products, questions and evaluation procedures. It does show why elegant prose cannot serve as a news-integrity test. An answer may have its main fact right yet omit a decisive perspective, assign an opinion to the wrong party or attach a link that does not support the relevant sentence.

The BBC/EBU toolkit sets a plain expectation: an assistant should provide sources for the key claims it makes (Yezza et al., 2025, p. 7). A block of links at the bottom is not sufficient. A user needs to know which item supports which claim, whether it is current, whether it contains the information attributed to it and whether it can be opened.

A source performs three democratic functions. It is evidence, allowing a claim to be checked. It is an address for accountability, leading to an author or institution that can answer a correction. It is also a window onto plurality, enabling the reader to leave the single synthesis. When an answer removes the source, it shortens not only the journey to information but also the journey to disagreement.

More than a quarrel over clicks

Publishers are right to worry that a no-click answer can absorb their reporting while removing the traffic that helps finance it. Without reporters, documents and local conversations, there is eventually little of value for an answer engine to summarise. The economic problem is real. It should not obscure the public one.

The larger stake is informational freedom and epistemic pluralism. Freedom of choice requires visible alternatives, especially when sources differ in interest, method or confidence. A single synthesis can be an excellent orientation device, but it cannot honestly present itself as a view from nowhere. Its order is the product of decisions.

The dignity of journalistic work is involved too. An original finding is not anonymous raw material equivalent to punctuation. Somebody obtained the document, travelled to the location, secured a witness’s consent and carried the publication risk. Visible authorship assigns credit and responsibility. When a model takes a finding, drops its maker and pronounces it in its own voice, public knowledge becomes detached from the process that produced it.

Nor do users have equal time and expertise. A specialist may treat the answer as a starting point; a newcomer is more likely to take it as the complete map. Research on news audiences records the growing role of platforms, video and intermediaries in access to news, alongside ambivalence towards personalisation (Newman et al., 2025, pp. 9–18, 48–52). Product design should not transfer the whole burden of audit to someone who cannot know what has been omitted.

There is a basic issue of dignity here. People affected by a protest, school closure or health policy deserve not to become ingredients in a frictionless paragraph whose origins nobody can inspect. A citizen’s capacity to object depends upon knowing whether a contested statement came from an official document, a reporter, a pressure group or the system’s own inference.

The Three Gates and a Trail protocol

DBMoJo proposes Three Gates and a Trail, a protocol for answers about news and public affairs. A provider can use it in product design; a newsroom can use it both to audit an external answer and to prepare its own material for accountable reuse. Every stage has an owner, a pass condition and a record.

1. Corpus gate: what could have been used? The provider states the broad source types, time range, access to the live web and material language or licensing limits. It need not reveal all training data to say whether this answer rests on current retrieval, a closed collection or model memory. Pass condition: a user can tell whether the system could possibly have encountered the newest document.

2. Retrieval gate: what was fetched now? Alongside the answer, show the specific items and, where possible, passages used for material claims. Record access dates and identify whether an item is primary, secondary, satirical or opinion. Pass condition: a reviewer can reach the passage that actually supports a sentence without going on a treasure hunt.

3. Synthesis gate: what did the system do to the material? Give claims an intelligible status: quotation, paraphrase, proposition corroborated across sources, inference, or unresolved. Disagreement must not vanish merely because a smooth paragraph is prettier. Pass condition: the audience can distinguish what the source stated from what the system inferred.

4. Accountability trail: where can a person return? Attach a timestamp and version, retain source authorship, provide an error-reporting route and record substantive corrections. In a newsroom the commissioning editor owns the trail; in a product, a named team owns news quality. Pass condition: when the answer changes, someone can establish what was corrected and why.

A publisher cannot control the gates of somebody else’s platform, but it can strengthen its own trail: unambiguous bylines, primary evidence next to the relevant claim, update dates, structured metadata, a compact ‘key evidence’ block and a public corrections record. None guarantees citation by an AI

system. This is not a new species of magical SEO. It is an accountable arrangement that also serves the human reader.

Dress rehearsal: has the school actually closed?

Suppose a resident asks whether the council has decided to close a school. A responsible answer should not open with a seamless yes or no if the documents show only a draft resolution. Under the protocol, the corpus gate might state that the system searched the official council bulletin and local reporting up to 6 p.m. The retrieval gate would display the draft resolution, the meeting agenda and a reporter’s account of the committee hearing.

The synthesis gate would separate the fact — the committee recommended the proposal — from the inference that the full council is likely to approve it. The accountability trail would give the generation time, preserve the journalists’ names and invite an update after the vote. This answer is slightly less magical because its screws are visible. That is precisely why it is more useful: the resident can decide whether the synthesis is enough or whether to read the resolution.

The protocol cannot solve everything. Displayed sources may themselves be one-sided; the corpus may have gaps; technical labels may overwhelm a hurried reader. The interface should therefore be layered: a concise answer, conspicuous sources and an expandable audit. Transparency does not mean emptying the engine onto the passenger seat. It means displaying the instruments required for safe travel.

There is also an irreducible editorial choice. No product can show everything, and no newsroom publishes the whole world. The purpose of the protocol is not total disclosure. It is the practical possibility of recognising selection, reaching evidence and disputing the hierarchy. That possibility must remain proportionate to the issue: a weather query needs less machinery than an allegation of electoral fraud.

An answer has a front page too

An answer engine does not abolish the agenda. It makes the agenda less visible. Paragraph order, source choice, tonal certainty and missing alternatives form a front page hidden inside the prose. The more complete an answer feels, the easier it is to forget that it could have been arranged differently.

Mature criticism does not demand that a system display everything. No medium ever has. It demands the ability to recognise selection, travel back to evidence and contest the hierarchy. One answer can be a convenient entrance to a subject. It should not become an emergency exit from pluralism.

The classic gatekeeper stood at a gate and at least wore a visible uniform. Today’s gates are written into infrastructure. We therefore need neither panic nor nostalgia for newsprint, but new trails of responsibility. When an answer has no visible front page, the duty to show its doors becomes more important, not less.