↖︎ Vishal Singh

The Political Ad Ledger · Chapter 5

What the Ads Say

The ledger records what was spent, when, and where — but never what an ad actually said. So we watched them. A close reading of the top video ads from five key advertisers reveals who builds a campaign around a leader, who attacks, and a government that puts the Prime Minister in most of its ads.

Everything in the first four chapters came from metadata — money, dates, formats, geography. The one thing the dataset never contains is the ad itself: its words, its images, its argument. To read those, you have to leave the table and watch the creative. This chapter does exactly that for a focused, high-value slice: the fifteen highest-spend video ads from each of five advertisers that anchor the earlier chapters — the BJP, the Congress, the Union government's ad bureau (CBC), and the two consultancies, I-PAC and Populus. Seventy-five ads in all, each watched end-to-end by a multimodal model applying a fixed codebook.

This is a close reading, not a survey: fifteen ads per advertiser is a portrait of each one's flagship messaging, not a population estimate. Read it that way. Even so, the portraits are sharp — and they line up with, and deepen, what the structured chapters implied.

1  The government's face

Chapter 4 showed the Union government's ad bureau is the second-largest political advertiser in the data. This chapter asks the question the metadata could not: what is in those ads? The answer is striking. In 87% of the CBC's top videos a specific politician appears, and in 53% a politician is the central subject — most often the Prime Minister. These are public-money advertisements, and more often than not they are built around a person.

How often each advertiser's ads are built around a politician

Share of the advertiser's top-15 videos in which a specific politician appears, and in which one is the central subject

Figure 1. The government (CBC, in red) is not an outlier for showing a leader — most advertisers do — but for a public body, centering one in a majority of ads is notable. The Congress is the mirror image: it shows leaders least and centers them in none of its top videos.

"The Modi government guarantees that no one will go hungry, providing free ration…"Central Bureau of Communication — a Government of India ad

The contrast with the Congress is the sharpest single finding here. The BJP shows the Prime Minister in 93% of its top ads; the Congress puts a leader at the center of none of its fifteen, leaning instead on promises and party. Two national parties, two opposite theories of what an ad is for.

2  Who attacks, who aspires

Coding each ad's dominant appeal — attack, contrast, achievement, aspiration, or neutral information — separates the advertisers cleanly. The government never attacks (its ads are pure achievement). The parties mix achievement and aspiration. And the consultancies are the most combative: I-PAC's dominant register is the attack ad.

The dominant appeal of each advertiser's top videos

Share of top-15 videos by primary appeal

Figure 2. Government (CBC) is entirely positive-achievement; parties blend achievement and aspiration; the consultancies carry the negativity. Hover a segment for the count.

"The BJP and its leaders disrespect Bengal's great personalities and culture…"I-PAC — a Trinamool Congress attack ad, West Bengal

This is the consultancy fingerprint made literal. I-PAC, working the West Bengal front, runs identity-and-grievance attack ads against the BJP. Populus, the DMK's in-house shop, runs the opposite register — leader-rally ads for M.K. Stalin — yet both build almost every ad around the Chief Minister (I-PAC centers a leader in 80% of its ads, Populus in 93%). The strategist's signature isn't a tone; it's a fixation on the candidate.

3  The beneficiary agenda

Across all seventy-five ads, one issue frame dominates: welfare schemes, named in 55% of the videos, followed by claims of past achievement (49%) and, strikingly, women as a target constituency (35%). This is the "labharthi" (beneficiary) politics that has reshaped Indian campaigning — free rations, cash transfers, bus-fare waivers, cooking-gas cylinders — rendered as advertising.

The issues these ads raise

Share of the 75 coded videos that raise each theme (ads raise several)

Figure 3. Welfare and record-touting lead; women appear as a named audience in over a third of ads. Multi-labelled, so shares sum past 100%.

"The Mudra scheme empowers women entrepreneurs, leading to self-reliance…"Bharatiya Janata Party

How much to trust this. These readings come from a single multimodal model applying a fixed codebook at temperature zero. An earlier pilot found that model coders agree strongly on observable, factual fields — whether an ad attacks, whom it depicts, what scheme it names — and less on subtler judgments of tone. Treat the "who appears / attack-or-not / which scheme" findings as solid, and the finer appeal distinctions as indicative. A human-coded gold standard and multi-model cross-check are the next step.

None of these portraits required the dataset to tell us what the ads said — it can't. But by resolving each ad to its video and watching it, the ledger's blind spot becomes readable. The picture that emerges is coherent with everything before it: a ruling party fused to its leader, an opposition still selling promises, a government advertising its own achievements with the Prime Minister's face, and a shadow layer of consultancies running the sharpest, most candidate-centered ads of all.

Data & methods

  • Which ads. The 15 highest-spend video creatives from each of BJP, INC, CBC/DAVP, I-PAC, and Populus — 75 ads. Chosen because these advertisers anchor Chapters 2–4 and their ads are disproportionately on YouTube (resolvable).
  • Acquisition. Each ad's Transparency Center page was rendered to recover its YouTube id; the video was downloaded and passed to the model. Descriptors resolved for 358 marquee videos; this chapter codes the top 15 per advertiser.
  • Coding. Gemini 2.5 Flash watched each video (audio + visual) and returned a structured record against a 21-field codebook (v1.1) at temperature zero — every response cached. It reads on-screen faces and text, so fields like "shows a politician" use the visual, not a transcript.
  • Honest limits. Fifteen ads per advertiser is a descriptive portrait, not a population estimate. Single-model coding; reliability was measured on a transcript pilot (strong on factual fields, weaker on tone) and a human gold standard is pending. Video only — image ads (the majority by count) are not yet coded.